Articles

[Updated on October 26th]Hello everyone, this time I will briefly explain his LoRA creation feature in TensorArt. This feature allows you to create your own learning files.⚠️As I am self-studying, I would appreciate it if you could read this for reference only⚠️overview:LoRA (low rank adaptation) is an efficient method for fine-tuning AI models. TensorArt allows you to easily create your own training files with specific styles and characteristics using LoRA. This allows you to generate images for individual projects and creative needs.process:1. Login and menu selection:First, log into TensorArt and select the Online Training option from the menu.2. Upload images:Add the images you want to train to the upload area on the left. To do this, prepare multiple images with specific characteristics, such as the character's expression or pose.3. Set model and trigger word:Select the model theme to be used, the base model (select from first, anime, reality, 2.5D, standard, custom), and set the trigger word etc.🌸 If you select Custom, you can now train using your favorite base model (SDXL, SD3, Hanyuan, FLUX).🌸Trigger words do not necessarily need to be set depending on the purpose.4. About tags:Tags are automatically generated for each image when you upload it. Click on the uploaded image to check, delete, or add the generated tags. By deleting the tags of the features you want to learn, you can learn the features more accurately. Adjusting the tags will improve the quality of the images produced.(If this is your first time, feel free to ignore it.)5. Run the training:Once the settings are complete, click the "Train Now" button at the bottom right. Training can take minutes to hours, and you can track your progress on a dedicated page. The amount of credits consumed will vary depending on the number of images you prepare, the number of training sessions, and the number of epochs, so please proceed in a planned manner.6. Download the file:Once training is complete, download the generated LoRA file and use it for actual image generation.7. Host the model:Proceed to create a project from "Host my model", enter the necessary information and click the create button.Completion and testing:Let's try image generation using the trained LoRA file. Although features can be learned sufficiently with a small number of images, similar compositions and poses are more likely to be generated if there are fewer reference images.　　　　　　　　　　　　　Completed: "Shizuku-chan" general-purpose XLCredit consumption:Creating LoRA consumes credits. The number of credits consumed varies depending on the number of images you prepare, the number of training sessions, and the number of epochs, so it is important to check the number of credits you need in advance and proceed accordingly. If you don't have enough credits, we recommend purchasing more or considering a Pro plan.summary:By using TensorArt's LoRA creation function, you can create illustrations with a higher degree of freedom. Please try out the features that are easy to use even for beginners. By using LoRA, your creative projects will be even more fulfilling.Let's create! !

First, let's open the online trainingThe first step is to upload the training set. Since the Kontext Dev model is a bit special, the training data needs to be in pairs each pair includes a “before editing” and “after editing” image. Upload the “before” images in the red section and the “after” ones in the green section.We recommend uploading around 10 to 20 pairs.Similarly, for Kontext Dev, what you need isn’t “image captions” you should enter simple editing instructions in the red box instead. Just a short phrase or a single sentence is enough. You can also add trigger words in the green box to help improve the LoRA’s performance after training.In the green box, fill in the preview settings for checking how your model will perform this includes uploading an original image and adding an editing prompt. That way, once training is done, you can easily see how well it worked.As for the settings, we recommend using the “Basic Mode” with the default values. But if you want more advanced control, you can switch to “Pro Mode” by clicking the red box in the top right corner.The "Start Image" is the one used to show comparison previews during each training epoch. Here, I uploaded an image that's part of the training set, but I actually recommend using one that’s not in the training set it gives you a better sense of how well the model generalizes.Once everything’s ready, just click “Start Training Now” to get things going.Remember, once the training is done and your model is published, make sure to use it in Image2Image mode here.

How to create Lora (Basic version)I believe that when creating an image you like, you may sometimes feel that some element is missing, and you may try to search for the relevant lora, but still cannot find it.At this time, you may think, should I make a lora myself? In this way, in my future works, maybe I can highlight more of the elements I like! I can also share it with my creative partners.Let’s get started!First, you need to collect more pictures of similar elements (about 12 pictures), and the picture resolution should be higher, so that when you use it with other LoRa in the future, the pictures will be clearer, unless the element you want is a hazy feeling.Next, it’s time to start training!First, open the user interface and there is a model I trained. Then click the online training in the upper left corner to enter the training interface. Add the prepared image in the lower left corner. Select the basic model type you often use in the upper right corner, such as: sd1.5, sd3.5, pony, flux, Hunyuan, etc.Next, you need to consider your computing power budget, because if you use different basic models to train the same image, the results will be very different, and the computing power consumption will also be different. Also, can you use it together with the commonly used LoRa?For the trigger word part, you can input the most important elements of this lora. You can choose not to input them first and wait until you see the training results.The prompt word part seems to affect the training results. If you have a clear goal, you can enter it. The default is 1girl. If your element is not a girl, you can change it or leave it blank.OK, next click on training, wait in line and check the training time.After the training is finished, I will look at the training results. There are ten training results in total. I usually choose one from the sixth to the tenth one that looks closer to the result I want and press publish.Select Create Project at the top.Enter the project name (be careful, this cannot be changed)Select Lora TypeAdd lora tagparameterI usually choose 500 for the number of iterations.Type the trigger word and descriptionSelecting a base modelReverse prompt wordUpload files unless they are the result of previous trainingOtherwise, I usually adjust the precision to fp32.Showcases (Image/Video)If you are uploading the results of a previous training, you need to upload the workbench image and cover imageOK, press PublishAdjust the details again, update and wait for the system to be deployed, then you can try it out and see your results!😁😁😁

My Journey: Training a LoRA Model for Game Art DesignWhat is LoRA?LoRA (Low-Rank Adaptation) is a powerful technique to create custom AI art models, perfect for game designers looking to develop unique visual styles.My Training Setup for Adrar Games Art StylePreparing Your Training DatasetTechnical SpecificationsBase Model: FLUX.1 - dev-fp8Training Approach: LoRA (Low-Rank Adaptation)Trigger Words: Adrr-GmzEpochs: 5Learning Rate: 0.0005 (UNet)Key Training ParametersNetwork ConfigurationDimension: 2Alpha: 16Optimizer: AdamW 8bitLR Scheduler: Cosine with RestartsAdvanced TechniquesNoise Offset: 0.1Multires Noise Discount: 0.1Multires Noise Iterations: 10Sample Prompt"A game art poster of a Hero standing in a fantastic ancient city in the background, and in the top a title in a bold stylized font 'Adrar Games'"My Learning ProcessChallengesCreating a consistent game art styleCapturing the essence of "Adrar Games" visual identityBalancing technical parameters with creative visionInsightsLoRA allows precise control over art generationCareful parameter tuning is crucialSmall adjustments can significantly impact resultsPractical TakeawaysStart with a clear artistic visionExperiment with different settingsDon't be afraid to iterate and refineRecommended Next StepsGenerate multiple sample imagesAnalyze and compare resultsAdjust parameters incrementallyBuild a library of unique game art assetsWould you like me to elaborate on any part of my LoRA training experience?

寒い冬は動きたくない。暑い夏は外に出たくない。花粉の多い春は家にいたい。枯れ葉が飛んでくる秋は窓から外を眺めたい。僕にとってGPTは代わりに色々してくれるいい相棒です。ご飯とかお風呂を用意してくれるようになったら完璧。そんな楽をすることしか考えてない僕の、ローラ詳細設定方法をご紹介します。ババン。GPTに詳細設定を考えてもらうSample画像を1枚アップロードしてこう聞いてください。赤字の部分を変更します。AI画像生成モデルのFlux用のLoRAを作る予定です。「結婚式場」をテーマにしたLoRAです。 添付画像のような画像を40枚用意しました。最適な「繰り返し数Repeat」「エポック数Epoch」「Network Rank Dim」「Network Alpha」などの設定を考えて貰えますか。タグ設定について注意点などありますか？サンプル画像生成用のプロンプトを作成して貰えますか。他にも重視したいことを記入すると良いですね。GPTさんの実際の返事LoRAの設定や学習手順について、結婚式場をテーマにしたLoRAを作成する際の最適な設定や注意点、サンプル生成プロンプトについて詳しく解説します。LoRA学習用の設定提供された40枚の画像を使用する場合、データ量は少なめですので、以下の設定を推奨します。基本設定Repeat（繰り返し数）：5～10データセットが40枚と少ないため、各画像を繰り返して学習データの量を補います。Epoch（エポック数）：10～15過学習を防ぎつつもデータ量に応じて十分に学習させるため、エポック数は10～15程度に調整します。Network Rank Dim（Rank）：16～32Rankを16～32程度に設定し、モデルの柔軟性を確保します。低すぎると学習内容が薄まり、高すぎると過学習のリスクが上がります。Network Alpha（Alpha）：16AlphaをRankと同値程度に設定すると、LoRAの学習が安定します。タグ設定の注意点共通タグ"wedding venue"、"flower decorations"、"arches"、"drapes"、"fairy lights" などの基本要素をタグとして追加。→ 結婚式場の要素がしっかり学習されるように意識します。詳細なタグ画像内の要素を細かく分けてタグ付け。例えば：花の種類：「roses」「floral arrangements」「bouquets」装飾：「candles」「lanterns」「string lights」「arched drapery」照明や雰囲気：「soft lighting」「romantic setting」「warm glow」固有名詞の避け方「特定の会場名」や「ブランド名」などの固有名詞は含めないよう注意。サンプル画像生成プロンプト学習が完了したLoRAを使って、以下のプロンプトでサンプル画像を生成できます。コードをコピーする(masterpiece, best quality, ultra-detailed), wedding venue, floral arch, draped curtains, fairy lights, romantic lighting, beautiful flower arrangements, glowing candles in glass lanterns, soft sunlight through tall windows, elegant decorations, roses, pastel colors, dreamy atmosphere, wooden floor, luxurious and enchanting wedding setting追加の強化要素季節感を出す場合：spring wedding や winter wedding theme背景変更：sunset lighting や nighttime fairy lights装飾の強調：detailed flower patterns や ornate arch design補足生成の際は、LoRAの強度（weight）を0.7～1.0程度で調整し、モデルの影響度を最適化します。サンプル生成後、LoRAが意図した要素をしっかり学習できているか、具体的に確認して調整を加えてください。この設定で試してみて、調整が必要な場合はフィードバックをお待ちしています！✨

A step-by-step guide on how to train a LoRA; part 3!Warning: This guide is based on Kohya_SSThis guide REQUIRES that you read "How I LoRA: A beginners guide to LoRA training | Part 1: Dataset Prep." and "How I LoRA: A beginners guide to LoRA training | Part 2: Training Basics"This guide CAN be ported to Tensor.art's trainer; if you know what you are doing.This guide is an (almost) 1:1 of the following guide: https://civitai.com/articles/3522/valstrixs-crash-course-guide-to-lora-and-lycoris-trainingEdits were made to keep it short and only dive into the crucial details. It also removes a lot of recommendations I DO NOT follow.; for more advanced information, please support the original guide. If you want to do things MY way, keep reading.THE SETTINGS USED ARE BASED ON SDXL, DO NOT FOLLOW IF YOU ARE TRAINING ON V-PRED OR 1.5Testing your LoRAThere are two ways to test a LoRA. During training and after training.During:While on your Kohya_ss, there is a section for a "test" prompt. Use it. If you followed the guide you should have set "save every N epoch" as 1. Meaning that every epoch it will save a model and by proxy, test it with the given prompt.Look at each image, and judge its quality.After (The right way):After training is done, move all your safetensors files to your lora folder on your WebUI instalation. I will asume you have A1111, A1111 Forge or A1111 Re-Forge (the best one).On your WebUI, set yourself up with all the settings you would normally use; checkpoint, scheduler, etc.Copy/paste one of your dataset prompts to the prompt area (This will test overfitting).Navigate to the LoRA subtab and add the first file; ie.: Shondo_Noob-000001.safetensor, this will add the LoRA to the prompt as: <lora:Shondo_Noob-000001:1>; change the :1 to :0.1Set a fixed seed; ie.: 1234567890Scroll down to the "script" area of your WebUI and select X/Y/ZSet your X, Y and Z as "Prompt S/R"On X; write all of your LoRA's filenames; ie.: Shondo_Noob-000001, Shondo_Noob-000002, Shondo_Noob-000003, Shondo_Noob-000004, etc. Depending on how many files you saved, their names, etc. ALWAYS SEPARATE WITH A COMMA.On Y; write all the strength variables from 0.1 to 1, ie.: 0.1, 0.2, 0.3, etc. ALWAYS SEPARATE WITH A COMMA.On Z; write an alternate tag to test flexibility, so, if your prompt is: "fallenshadow, standing, dress, smile", write something like: dress, nude, swimwear, underwear, etc. This will create a grid where instead of wearing a dress, she will be nude, wear a swimsuit, etc. ALWAYS SEPARATE WITH A COMMA.If you did a concept LoRA or a style lora:On your WebUI, set yourself up with all the settings you would normally use; checkpoint, scheduler, etc.Copy/paste one of your dataset prompts to the prompt area (This will test overfitting).Navigate to the LoRA subtab and add the first file; ie.: doggystyle-000001.safetensor, this will add the LoRA to the prompt as: <lora:doggystyle-000001:1>; change the :1 to :0.1Set a fixed seed; ie.: 1234567890Scroll down to the "script" area of your WebUI and select X/Y/ZSet your X and Y as "Prompt S/R"On X; write all of your LoRA's filenames; ie.: doggystyle-000001, doggystyle-000002, Shondo_Noob-000003, doggystyle-000004, etc. Depending on how many files you saved, their names, etc. ALWAYS SEPARATE WITH A COMMA.On Y; write all the strength variables from 0.1 to 1, ie.: 0.1, 0.2, 0.3, etc. ALWAYS SEPARATE WITH A COMMA.Selecting the right fileOnce the process finishes, you should have at least 2 grids, one XY with dress, and another with nude (for example). Or one if you didnt set up an Z grid. Up to you.Now look at the grid and look for the "best" result. Look at the art style bias, pose bias, look bias, etc. The more flexible the better. If on fallenshadow-000005 shondo's pose is always unique but after 000006 she's always standing the same way, ignore 000006+If at some point the art style gets ignored or changes and fixates on it; ignore it.If at some point ANYTHING starts repeating that you don't want; ignore it.The only thing that should repeat at all times is whatever corresponds to the trained concept. If you only trained a wolf with a hat but it should always be a different hat, avoid a file that gives him the same hat on the same pose with the same style.If the result image is identical to the training data; avoid it! You are not here to do the same images as your data, you are here to make new ones, remember?If colors are weird; bad.If shapes are mushy; bad.If angle is always the same; bad (unless you prompted for it).Anything that goes against the concept or the flexibility of it: BAD.Any file that has to be lower than 1 or 0.9: BAD. If your LoRA "works best" at 0.6 strenght, it's shit.THIS IS IT FOR PART 3. Now do some good cool loras.

Let's talk about Illustrious and NoobAI LoRA'sPrefaceI am currently using tensor.art with Professional Mode to train my Lora, this article will mainly discuss what I've tried and I welcome others to discuss too as there's no official finetune guide.GuidelinesHigher rates = stronger character features but potential loss in image qualityLower rates = better image quality but weaker character featuresMost character Loras work well with UNET around 0.0003 and TE around 0.00003Lower learning rates will adapt the features better but can also take longer. As for the dataset lets say i have 40 images , 5-10 repeats, 10 epochs, 4 batch size, this usually adds up to the total steps and then hopefully a model is trained well enoughThe ideal ratio is typically UNET:TE = 10:1UNET Rates (0.0005 - 0.0001):0.0005: Very strong influence, can overpower the base model. Good for exact character matching but may reduce image quality0.0003: Balanced influence, commonly used for character Loras0.0001: Subtle influence, maintains high image quality but character features may be less pronouncedText Encoder (TE) Rates (0.00005 - 0.00001):0.00005: Strong text conditioning, helps with character recognition0.00003: Moderate text influence, good balance for most character Loras0.00001: Light text conditioning, useful when you want minimal style transferDimension Ranks (DR) - Network Dim32: Standard/Default rank, good balance of detail and file size64: Higher detail capture, larger file size128: Very high detail, much larger file size256: Maximum detail, extremely large file sizeNetwork Alpha (AR) - Network AlphaAlpha is typically set to match or be slightly lower & higher than the rank.Common ratios:AR may be half the rank or even a quarter less than the DRAR: Standard training stability (1:1 ratio), same as the DRAR× 1.5: Increased stability, a quarter more than the DRAR× 2: Maximum stability, double the DRThe values below are not 100% but they are being figured out still.Basic Character Lora (Base Model's preference)DR 64, AR 32 - Best for: Simple anime/cartoon characters - File size: ~70MB - Good balance of detail and stabilityComplex Character LoraDR 64-48, AR 32-24 - Best for: Most character types - File size: ~100MB - Excellent for anime/game charactersStyle Loraexample : https://tensor.art/models/806682226684073145/NAI3-Kawaii-Style-Illustrious-NoobAI-nai-IL-V0.1example : https://tensor.art/models/806356844256811271/Anima-Crayon-Sketch-Illustrious-IL-V0.1original article says : DR 128, AR 64 to 32 - seems to be the best for a combination of complex features etc if the style is very detailed. otherwise lower ranks work too.Learning rates can vary: CAME and RAWR = 0.0002 UNET and 0.00002 TE will need about 2500 to 3000 steps ADAMW8BIT & ADAFACTOR between 0.0003-0.0005 UNET and 0.00003-0.00005 at 1000 steps but what i use instead :Parameter Settings Network Module LoRA Use Base Model rMix NNNoobAI - V1.1 Trigger words nai3_kawaii Image Processing Parameters Repeat 10 Epoch 10 Save Every N Epochs 1 Training Parameters Seed - Clip Skip - Text Encoder learning rate 0.00004 Unet learning rate 0.00035 LR Scheduler cosine_with_restarts Optimizer AdamW8bit Network Dim 32 Network Alpha 16 Gradient Accumulation Steps - Label Parameters Shuffle caption true Keep n tokens 1 Advanced Parameters Noise offset 0.0357 Multires noise discount 0.15 Multires noise iterations 8 conv_dim - conv_alpha - Batch Size 2 Sample Image Settings Prompt nai3_kawaii 1girl solo long hair looking at viewer blush bangs blue eyes hair ornament dress ribbon sitting closed mouth pink hair sleeveless hairclip sailor collar two side up book blue dress sailor dress . masterpiece, best quality, amazing quality, very aesthetic, absurdres Sampler eulerWhat works?I'd like to hear what works and doesn't work for illustrious:OptimizerLearning Rates could change dependent on the optimizer chosen.SchedulerNetwork Settings(DR) Dimension rank 128, 96, 64, 32, 16, 4(AR) Alpha rank 128, 96, 64, 32, 16, 4Don't use:ProdigyCan use:AdamW8BitConstant0.0003 LR (TE & UNET) - Aggressive Learning for characters0.0002 LR - Medium learning for characters (DR 128 AR 64)AdaFactorSchedulerCosine with restart0.0005-0.0003 LR (UNET)0.00005-0.00003 LR (TE)DR 128-32, AR 64-16 - usually i go half the Network Dimension Rankplagiarized and inspired from : https://civitai.com/articles/9148/illustrious-lora-training-discussionmodel used for my training : rMix NNNoobAI v1.1 - https://tensor.art/models/805164110363975687

link to Lora: https://tensor.art/models/806808906744706431/Podoru-Meme-LoRa-e10I fetched my training images from: https://apps.apple.com/us/app/sankaku-anime-ai-girlfriend/Pic collage shown here was created using https://gandr.io/Then I used the built in Tensor Art caption tool to set the promptsExample:"padoru 1girl solo long hair smile open mouth simple background hair ornament long sleeves hat white background holding tail full body yellow eyes :d red hair pointy ears hairclip chibi fur trim capelet fangs transparent background blush stickers monster girl slit pupils christmas red headwear santa hat santa costume meme scales sack lamia holding sack miia \(monster musume\)"Training settings:base model: NovaXL (as shown as base model in LoRa) Unet learning rate: 0.0001Repeat : 20 , Epochs 10Constant learning rate Network dim: 64 , Alpha: 32Loss chart is absurdly low:And here are some examples of the output from the NovaXL illustrious model:Very cute!Thats all for this training guide. Merry Christmas! /Adcom

EnglishToday, Ｉwill teach you how to use TensorArt to train an Hunyuan model online.When you have finished reading this tutorial and have trained a model, you can test the best raw image Step and CFG of your trained model in this workflow.👉https://tensor.art/workflows/877479664430371272Step 1: Open “Online Training.On the left side, you will see the dataset window, which is empty by default. You can upload some images to create a dataset or upload a dataset zip file. The zip file can include annotation files, following the same format as kohya-ss, where each image file corresponds to a text annotation file with the same name.In the model theme section on the right, you can choose from options such as anime characters, real people, 2.5D, standard, and custom.Here, we select “Base” and choose the Hunyuan model as the base model.For the base model parameter settings, we recommend setting the number of repetitions per image to 4 and the number of epochs to 16.、After uploading a processed dataset, if your dataset annotations include character names, you don’t need to specify a trigger word. Otherwise, you should assign a simple trigger word to your model, such as a character name or style name.Next, select an annotation file from the dataset to use as a preview prompt.If you want to use Professional Mode, click the button in the top right corner to switch to Professional Mode.In Professional Mode, it is recommended to double the learning rateand use the cosine_with_restarts learning rate scheduler. For the optimizer, you can choose AdamW8bit.Enable label shuffling and ensure that the first token remains unchanged (especially if you have a character name trigger word as the first token).Disable the noise offset feature, and you can set the convolution DIM to 8 and Alpha to 1.In the sample settings, add the Negative prompts, and then you can start the training process.In the training queue, you can view the current loss value chart and the four sample images generated for each epoch.Finally, you can choose the epoch with the best results to download to your local machine or publish directly on TensorArt.After a few minutes, your model will be deployed and ready.日本語今日、私はTensorArtを使用してHunyuanモデルをオンラインでトレーニングする方法を教えます。ステップ1: 「オンライントレーニング」を開きます。左側にデータセットウィンドウが表示され、デフォルトでは空です。データセットを作成するために画像をアップロードするか、データセットのzipファイルをアップロードできます。zipファイルには、kohya-ssと同じ形式のアノテーションファイルを含めることができ、各画像ファイルには同じ名前のテキストアノテーションファイルが対応しています。右側のモデルテーマセクションでは、アニメキャラクター、実在の人物、2.5D、標準、カスタムなどのオプションから選択できます。ここでは「Base」を選択し、Hunyuanモデルをベースモデルとして選びます。ベースモデルのパラメーター設定では、画像ごとの繰り返し回数を4、エポック数を16に設定することをお勧めします。 処理済みのデータセットをアップロードした後、データセットのアノテーションにキャラクター名が含まれている場合は、トリガーワードを指定する必要はありません。それ以外の場合は、キャラクター名やスタイル名など、モデルに簡単なトリガーワードを割り当ててください。 次に、プレビュー用プロンプトとして使用するために、データセットからアノテーションファイルを選択します。プロフェッショナルモードを使用したい場合は、右上隅のボタンをクリックしてプロフェッショナルモードに切り替えます。プロフェッショナルモードでは、学習率を倍増することをお勧めします。また、cosine_with_restarts学習率スケジューラーを使用してください。オプティマイザーとしては、AdamW8bitを選択できます。ラベルシャッフルを有効にし、最初のトークンが変更されないようにします（特にキャラクター名トリガーワードが最初のトークンの場合）。ノイズオフセット機能を無効にし、畳み込みDIMを8、Alphaを1に設定できます。サンプル設定でNegative promptsを追加し、その後、トレーニングプロセスを開始できます。トレーニングキューでは、現在の損失値チャートと各エポックごとに生成された4つのサンプル画像を表示できます。最後に、最良の結果が得られたエポックを選択して、ローカルマシンにダウンロードするか、直接TensorArtで公開できます。数分後には、モデルがデプロイされ、使用可能になります。한국인오늘은 TensorArt를 사용하여 Hunyuan 모델을 온라인에서 훈련하는 방법을 알려드리겠습니다.1단계: “온라인 훈련”을 엽니다.왼쪽에는 기본적으로 비어 있는 데이터셋 창이 표시됩니다. 데이터셋을 만들기 위해 이미지를 업로드하거나 데이터셋 zip 파일을 업로드할 수 있습니다. zip 파일에는 kohya-ss와 같은 형식의 주석 파일이 포함될 수 있으며, 각 이미지 파일에는 동일한 이름의 텍스트 주석 파일이 대응됩니다.오른쪽의 모델 테마 섹션에서는 애니메이션 캐릭터, 실제 인물, 2.5D, 표준, 사용자 정의 등 다양한 옵션 중에서 선택할 수 있습니다.여기에서는 “Base”를 선택하고 Hunyuan 모델을 기본 모델로 선택합니다.기본 모델 파라미터 설정에서는 이미지당 반복 횟수를 4로, 에포크 수를 16으로 설정하는 것을 권장합니다. 처리된 데이터셋을 업로드한 후, 데이터셋의 주석에 캐릭터 이름이 포함되어 있으면 트리거 단어를 지정할 필요가 없습니다. 그렇지 않으면 모델에 간단한 트리거 단어를 지정해야 합니다, 예를 들어 캐릭터 이름이나 스타일 이름 등. 다음으로, 미리 보기 프롬프트로 사용할 주석 파일을 데이터셋에서 선택합니다.전문 모드를 사용하려면, 오른쪽 상단의 버튼을 클릭하여 전문 모드로 전환합니다.전문 모드에서는 학습률을 두 배로 늘리는 것이 좋습니다.또한 cosine_with_restarts 학습률 스케줄러를 사용합니다. 옵티마이저로는 AdamW8bit을 선택할 수 있습니다.레이블 셔플을 활성화하고 첫 번째 토큰이 변경되지 않도록 합니다(특히 캐릭터 이름 트리거 단어가 첫 번째 토큰인 경우).노이즈 오프셋 기능을 비활성화하고, 컨볼루션 DIM을 8로, Alpha를 1로 설정할 수 있습니다.샘플 설정에서 Negative prompts를 추가한 후, 훈련 프로세스를 시작할 수 있습니다.훈련 대기열에서 현재 손실 값 차트와 각 에포크에 대해 생성된 4개의 샘플 이미지를 볼 수 있습니다.마지막으로, 가장 좋은 결과를 얻은 에포크를 선택하여 로컬 컴퓨터로 다운로드하거나 직접 TensorArt에 게시할 수 있습니다.몇 분 후, 모델이 배포되고 사용 가능해집니다.Tiếng ViệtHôm nay, tôi sẽ hướng dẫn bạn cách sử dụng TensorArt để đào tạo mô hình Hunyuan trực tuyến.Bước 1: Mở “Đào tạo trực tuyến.”Ở bên trái, bạn sẽ thấy cửa sổ tập dữ liệu, mặc định là trống. Bạn có thể tải lên một số hình ảnh để tạo tập dữ liệu hoặc tải lên tệp zip của tập dữ liệu. Tệp zip có thể bao gồm các tệp chú thích, theo cùng một định dạng như kohya-ss, trong đó mỗi tệp hình ảnh tương ứng với một tệp chú thích văn bản cùng tên.Ở phần chủ đề mô hình bên phải, bạn có thể chọn từ các tùy chọn như nhân vật anime, người thật, 2.5D, tiêu chuẩn và tùy chỉnh.Tại đây, chúng ta chọn “Base” và chọn mô hình Hunyuan làm mô hình cơ bản.Đối với cài đặt tham số của mô hình cơ bản, chúng tôi khuyên bạn nên đặt số lần lặp lại trên mỗi hình ảnh là 4 và số epoch là 16. Sau khi tải lên tập dữ liệu đã xử lý, nếu các chú thích của tập dữ liệu của bạn bao gồm tên nhân vật, bạn không cần phải chỉ định từ kích hoạt. Ngược lại, bạn nên gán một từ kích hoạt đơn giản cho mô hình của mình, chẳng hạn như tên nhân vật hoặc tên phong cách. Tiếp theo, chọn một tệp chú thích từ tập dữ liệu để sử dụng làm lời nhắc xem trước.Nếu bạn muốn sử dụng Chế độ Chuyên nghiệp, hãy nhấp vào nút ở góc trên bên phải để chuyển sang Chế độ Chuyên nghiệp.Trong Chế độ Chuyên nghiệp, nên gấp đôi tỷ lệ học.Và sử dụng bộ lập lịch tỷ lệ học cosine_with_restarts. Đối với bộ tối ưu hóa, bạn có thể chọn AdamW8bit.Kích hoạt xáo trộn nhãn và đảm bảo rằng mã thông báo đầu tiên không thay đổi (đặc biệt nếu bạn có từ kích hoạt tên nhân vật là mã thông báo đầu tiên).Tắt tính năng dịch chuyển tiếng ồn và bạn có thể đặt DIM tích chập là 8 và Alpha là 1.Trong cài đặt mẫu, thêm các Lời nhắc tiêu cực, sau đó bạn có thể bắt đầu quá trình đào tạo.Trong hàng đợi đào tạo, bạn có thể xem biểu đồ giá trị tổn thất hiện tại và bốn hình ảnh mẫu được tạo ra cho mỗi epoch.Cuối cùng, bạn có thể chọn epoch có kết quả tốt nhất để tải xuống máy tính của bạn hoặc xuất bản trực tiếp trên TensorArt.Sau vài phút, mô hình của bạn sẽ được triển khai và sẵn sàng sử dụng.españolHoy, te enseñaré cómo usar TensorArt para entrenar un modelo Hunyuan en línea.Paso 1: Abre “Entrenamiento en línea.”A la izquierda, verás la ventana del conjunto de datos, que está vacía por defecto. Puedes subir algunas imágenes para crear un conjunto de datos o subir un archivo zip del conjunto de datos. El archivo zip puede incluir archivos de anotación, siguiendo el mismo formato que kohya-ss, donde cada archivo de imagen corresponde a un archivo de anotación de texto con el mismo nombre.En la sección de temas del modelo a la derecha, puedes elegir entre opciones como personajes de anime, personas reales, 2.5D, estándar y personalizado.Aquí, seleccionamos “Base” y elegimos el modelo Hunyuan como el modelo base.Para la configuración de parámetros del modelo base, te recomendamos configurar el número de repeticiones por imagen a 4 y el número de épocas a 16. Después de subir un conjunto de datos procesado, si las anotaciones de tu conjunto de datos incluyen nombres de personajes, no necesitas especificar una palabra de activación. De lo contrario, deberías asignar una palabra de activación simple a tu modelo, como un nombre de personaje o un nombre de estilo. A continuación, selecciona un archivo de anotación del conjunto de datos para usarlo como un aviso de vista previa.Si deseas usar el Modo Profesional, haz clic en el botón en la esquina superior derecha para cambiar al Modo Profesional.En el Modo Profesional, se recomienda duplicar la tasa de aprendizaje.Y usar el programador de tasa de aprendizaje cosine_with_restarts. Para el optimizador, puedes elegir AdamW8bit.Habilita el barajado de etiquetas y asegúrate de que el primer token permanezca sin cambios (especialmente si tienes una palabra de activación de nombre de personaje como el primer token).Desactiva la función de desplazamiento de ruido y puedes configurar el DIM de convolución a 8 y Alpha a 1.En la configuración de muestra, añade los Avisos Negativos, y luego puedes comenzar el proceso de entrenamiento.En la cola de entrenamiento, puedes ver el gráfico del valor de pérdida actual y las cuatro imágenes de muestra generadas para cada época.Finalmente, puedes elegir la época con los mejores resultados para descargarla a tu máquina local o publicarla directamente en TensorArt.Después de unos minutos, tu modelo estará desplegado y listo para usar.

In this guide, we’ll walk through the full process of online training on TensorArt using Wan2.2. For this demo, we’ll be using image2video training so you can see direct results.Step 1 – Open Online TrainingGo to the Online Training page.Here, you can choose between Text2Video or Image2Video.👉 For this tutorial, we’ll select Image2Video.Step 2 – Upload Training DataUpload the materials you want to train on.You can upload them one by one.Or, if you’ve prepared everything locally, just zip the files and upload the package.Step 3 – Adjust ParametersOnce the data is uploaded, you’ll see the parameter panel on the right.💡 Tip: If you’re training with video clips, keep them around 5 seconds for the best results.Step 4 – Set Prompts & Preview FramesThe prompt field defines what kind of results you’ll see during and after training.As training progresses, you’ll see epoch previews. This helps you decide which version of the model looks best.For image-to-video LoRA training, you can also set the first frame of the preview video.Step 5 – Start TrainingClick Start Training once your setup is ready.When training completes, each epoch will generate a preview video.You can then review these previews and publish the epoch that delivers the best result.Step 6 – Publish Your ModelAfter publishing, wait a few minutes and your Wan2.2 LoRA model will be ready to use.Recommended Training Parameters (Balanced Quality)Network Module: LoRABase Model: Wan2.2 – i2v-high-noise-a14bTrigger words: (use a unique short tag, e.g. your_project_tag*)*Image Processing ParametersRepeat: 1Epoch: 12Save Every N Epochs: 1–2Video Processing ParametersFrame Samples: 16Target Frames: 20Training ParametersSeed: –Clip Skip: –Text Encoder LR: 1e-5UNet LR: 8e-5 (lower than 1e-4 for more stability)LR Scheduler: cosine (warmup 100 steps if available)Optimizer: AdamW8bitNetwork Dim: 64Network Alpha: 32Gradient Accumulation Steps: 2 (use 1 if VRAM is limited)Label ParametersShuffle caption: –Keep n tokens: –Advanced ParametersNoise offset: 0.025–0.03 (recommended 0.03)Multires noise discount: 0.1Multires noise iterations: 10conv_dim: –conv_alpha: –Batch Size: 1–2 (depending on VRAM)Video Length: 2Sample Image SettingsSampler: eulerPrompt (example):TipsKeep training videos around ~5 seconds for best results.Use a consistent dataset (lighting, framing, style) to avoid drift.If previews show overfitting (blurry details, jitter), lower UNet LR to 6e-5 or reduce Epochs to 10.For stronger style binding: increase Network Dim → 96 and Alpha → 64, while lowering UNet LR → 6e-5.

Wan2.2 and Qwen-Image Challenge is about to begin!Come and learn how to train LoRAs for these two models.First, open Online TrainingClick Standard and select Qwen-Image as the base model.What excites creators is the LoRA fine-tuning technique—with just 10 images, you can teach Qwen-Image your own unique style.Step 1: Prepare your training dataset in 10 minutesDataset: 10–50 images with a consistent style, theme, or subject. Image size is not restricted.Trigger word: Define a custom keyword for your style (e.g., “fashion_style”). Later, when generating images, you can use this word to apply the style. Set the number of repeats per image to 20, training epochs to 10, and fill in the LoRA model name along with other related parameters.Enter the prompt for preview generation in the Model Effect Preview Prompt input box. Qwen-Image supports prompts in Chinese.Here’s the example I wrote:“Create a humorous promotional poster featuring a cat wearing sunglasses, illustrated in a white-outlined cutout style, showing both a confused and cool expression. The background should be bright yellow with a folded texture. At the top, place a bold English title ‘STAY COOL’, and at the bottom add smaller Korean text. Include comic-style exclamation marks, arrows, and hand-drawn effects. The overall look should be quirky yet fashionable.”Once everything is set, click Start Training Now.During training, preview images will help you decide which LoRA performs best.Select the best LoRA and publish it.After waiting a few minutes for deployment, you can start running it.

Z- image Online Training - First Time FREE! & 50% OFFBy anchoring in the SFT stage, Z-Image achieves a superior balance between visual fidelity, generative diversity, and precise instruction following.It's your go-to choice due to its rapid training speeds and minimal dataset requirements.🔥 Limited Time OfferTo celebrate our launch and empower creators, we are offering50% OFF on all online trainingyour FIRST session is on us!Note: We will rebate up to 1,000 credits to your account upon the successful completion of your first training.Don’t miss out!📌 Go to Training 👉 https://tensor.art/trainTraining Tutorial1. Online Training WorkbenchIn the online training workbench, select Custom as the model type, then choose z-image as the base foundation model for training.2. Training SamplesSelect high-quality samples according to the training task. The recommended number of samples is approximately 30–300, and the more, the better.In our internal testing, the Base model uses a relatively high proportion of photorealistic style data during the SFT (Supervised Fine-Tuning) stage, which results in excellent image quality for realistic styles.3. Data Processing & LabelingThe workbench provides multiple advanced multimodal models for Labeling, including the latest Gemini-3-Flash model.You can use the default caption prompt for labeling, or customize your own prompt. After annotation is completed, add a trigger word and a training sample preview prompt. For example, jzxdda_style can be used as a trigger token — it carries no inherent semantic meaning and is used solely to activate the learned style.4. Start TrainingClick the “Start Training” button to launch the training task.The training details page will display the estimated remaining time, which typically ranges from 10 minutes to 1 hour, depending on the number of training samples and training steps.During training, you can monitor real-time loss curves and accuracy changes. The task can be stopped early at an appropriate epoch to prevent overfitting.5. Testing & PublishingClick “Publish” to create a project and deploy the trained model.

This is an easy step-by-step method to create your own LoRA using the base models available on Tensor.ArtOpen the Tensor.Art website (obviously, if you’re reading this article, you’re already on the site — lol).Click on your profile icon and select Training.Choose Online Training.Upload your sample images.Select the model type you want to train.Set the credits you want to use (minimum 10 credits).Choose the number of training steps / iterations for the model(I usually set this to around 20–30 steps).Enter the keywords you want the model to generate, such as 1girl, 1boy,or any other keywords you want to emphasize.Click Start to begin the process.Select an image style from the trained model.In most cases, the result will already be about 70-80% similar to the original,depending on the trigger words you use.Don’t focus too much on matching the sample images perfectly.Just choose a style that’s close to what you want, because later onyou’ll be adjusting parameters and mixing it with a base model during actual use anyway.Click Create Project.Click Confirm.Enter the name you want.Select a category.Add tags.Click Create.Enter the version name you want.Enter the trigger words to use the model.Add a short description.Set the minimum parameter (recommended: 0.5).No need to upload additional files.Select the image you want to display (you can upload more images if you want).Click Create.

The Importance of Proper Dataset Selection in Training to Prevent OverfittingIn the realm of machine learning, achieving a well-performing model hinges significantly on the quality and appropriateness of the training dataset. One of the critical challenges faced during model training is overfitting, where the model learns the training data too well, including its noise and outliers, resulting in poor generalization to new, unseen data. To mitigate overfitting, it's imperative to select and curate the right dataset. Here's why a proper dataset is essential in preventing overfitting and how it can be achieved.Understanding OverfittingOverfitting occurs when a model becomes overly complex, capturing not only the underlying patterns in the training data but also the noise. This leads to high accuracy on the training dataset but poor performance on validation or test datasets. Essentially, an overfitted model has memorized the training data rather than learning to generalize from it. This issue is particularly prevalent in datasets that are too small, noisy, or unrepresentative of the problem space.The Role of a Proper DatasetDiversity and Representativeness: A good dataset should be diverse and representative of the various scenarios the model will encounter in real-world applications. This means including a wide range of examples, ensuring that the model learns to generalize from different patterns and conditions rather than memorizing specific instances.Sufficient Size: The size of the dataset is a crucial factor. Small datasets often lead to overfitting because the model doesn't have enough examples to learn the underlying patterns adequately. Larger datasets provide more opportunities for the model to see varied examples, reducing the chance of overfitting.Balanced and Unbiased Data: An imbalanced dataset, where certain classes or conditions are overrepresented, can cause the model to be biased towards those classes. This imbalance leads to overfitting on the overrepresented classes. Ensuring that the dataset is balanced helps the model learn to generalize across all classes more effectively.Clean and Preprocessed Data: Noisy data with errors or irrelevant information can mislead the model during training. Proper preprocessing, such as removing outliers, normalizing values, and handling missing data, is essential to provide the model with clean data that accurately reflects the problem domain.Augmentation Techniques: Data augmentation involves creating variations of the training data through transformations such as rotations, translations, and scaling. This technique artificially increases the dataset size and diversity, helping to prevent overfitting by exposing the model to more varied examples.Strategies to Ensure a Proper DatasetCross-Validation: Using cross-validation techniques, where the dataset is split into multiple training and validation sets, can provide a better estimate of the model's performance and help in identifying overfitting. This method ensures that the model is tested on different subsets of data, promoting better generalization.Regularization: Applying regularization techniques such as L1 or L2 regularization can help in penalizing overly complex models, encouraging simpler models that generalize better. This approach works well in conjunction with a well-curated dataset to prevent overfitting.Data Splitting: Properly splitting the data into training, validation, and test sets is crucial. The training set should be used to train the model, the validation set to tune hyperparameters, and the test set to evaluate the final model performance. Ensuring that these sets are representative of the entire dataset helps in achieving a balanced training process.Monitoring Learning Curves: By monitoring the learning curves of training and validation losses, practitioners can identify signs of overfitting early. If the training loss continues to decrease while the validation loss starts increasing, it's a clear indication of overfitting.

1.Adding and processing datasetsClick "Online Training" on the Tensorart homepage to enter 1.1 Add dataset1.1.1Dataset l Currently supported formats are png/jpg/jpeg, and up to 1000 images can be added for training.l The uploaded picture can be deleted by clicking on the upper right corner.l It is recommended to upload higher definition images as much as possible for better training resultsl Enhanced datasets can be added, such as cropping and segmentation, image mirroring/flipping 1.1.2Regularized datasetl Regularization technology is widely used in Machine Learning and Deep learning algorithms. Its essential function is to regularize and reduce the weight of training materials, prevent overfitting, and improve Model Generalization Ability.l We can upload regularized datasets here, and the regularized dataset can be generated using the base model used for training.l For pure beginners who are completely unfamiliar with the training process, not using regular expression datasets may achieve better results.Please do not upload any illegal images such as bloody/violent/yellow/political images. Uploading illegal images multiple times may result in account suspensiorn 1.2 Batch clipping  1. Cutting method:Focus cropping: crop according to the main content of the picture. Center Crop: Crop the central part of the picture2. Choose the cutting size according to the training bottom filmSD1.5 optional sizes:n 512x468n 512x512n 768x512SDXL optional sizes:n 768x1024n 1024x1024n 1024x7681.3 Automatic markingEach uploaded image will be automatically tagged, and thetag content can be viewed by clicking on the image. In addition, you can also add and delete image tags. 1. If the training character wants to fix certain features, the procompt word for that feature can bedeleted2.Any AI automatic labeling cannot be 100% accurate. Ifpossible, manually screen once to eliminate incorrect labeling and improve the quality ofthe mode  1.4 Batch labelingCurrently, it supports adding tags to images in batches. You can choose to add them to the beginning or end of the line. Generally, they are added to the first line as a trigger word. 2.Training parameter settings 2.1 Set the number of repetitionsThe number of repetitions of image training, that is, the repeat parameter. Generally speaking, when training locally, this parameter needs to be adjusted separately in the training dataset folder In the online training workbench of Tensorart, we can change thenumber of repetitions of individual images from here. If you upload an enhanced dataset on it, you can sset different repetitions here 2.2 Base modelModel theme & base model selection: The model presets different training parameters according to different themes. Choosing the appropriate base model will make your Model Training twice as effective! Note: LoRA between different XL models is unlikely to be universal, please choose the base model carefully Two-dimensional characters: Optional base model AnythingV5/Animagine XL/Kohaku-XL Delta, training SD1.5 two- dimensional character LoRA requires AnythingV5, training SDXL LORA requires Animagine XL/Kohaku-XL Real person: Optional base model EpiCRealism (SD 1.5)/Juggernaut XL (SDXL). Some parameters of the training haVe been preset, and you can adjust the relevant parameters according to your needs 2.5D: Optional base model DreamShaper/GuoFeng3/DreamShapper XL1.0/GuoFeng4 XL, training SD1.5 LoRA requires DreamShaper/GuoFeng3, training SDXL LORA requires DreamShaper XL1.0/GuoFeng4 XL Standard: Default use of SDXL1.0/SD1.5 base as the training model, if not special needs, it is not recommended to useSingle Repeat (Repeat):Repeat refers to the number oftimes AI learns for each image, where Repeat only takes effect for imagesthat have not been set separatelyTraining rounds (Epoch): Epoch refers to a cycle of AI learnirng from your images. After all the images have completed Repeat, tthis is an Epoch. Total Steps: see the supplement below the table Model Effect Preview Prompt Word: The prompt word Ihere is a preset image for each version saved by Epoch, used to preview tthe training effect of the model This parameter does not affect the training effect and the qualitof the model, and is only used as a real-time preview graph parameter The formula for calculating the total number of steps is:(Number of images in the training dataset Repeat Epoch)The total number of steps directly affects the computing power con:sumption of Model Training, and the more steps, the greater the computing power consumption 2.3 Professional modeIt is not recommended for beginners to use professional mode Number of repetitions per image (Repeat): Repeat refers to the number of times AI learns from each imageTraining rounds (Epoch): Epoch refers to a cycle of AI learning from your images. After all the images have completed Repeat, this is an Epoch. The formula for calculating the total number ofsteps is:Number of images in the training dataset * Repeat* EpochThe total number of steps will directly affect the computing power consumption of Model Training. The more steps, the greater the computing power consumption Seed (seed): (metaphysics, random can be)Text Encoder Learning Rate : Adjust the sensitivity of the entire model to tags If you find unnecessary items during the image generation process, you need to reduce the TE Learning Rate; if you find it difficult to make the content appear without heavily weighting the prompts, you need to increase the TE Learning Rate. Unet Learning Rate: the speed and degree of model learning High Learning Rate can make AI learn faster, but may lead to Overfitting. If the model cannot replicate the details and the generated graph does not look alike, then the Learning Rate is too low. Try increasing the Learning Rate Learning Rate Scheduler: Scheduler refers to "how to set the change of Learning Rate" Optimizer: The optimizer is set up to update the weights of neural networks during training. Various methods have been proposed for intelligent training. Training grid size Dim: DIM represents the dimension of neural networks. The larger the dimension, the stronger the expression ability of the model, and the larger the final volume of the model. DIM is not the bigger the better. For a single character LORA, there is no need to open 128 for DIM. Training Network Alpha Value:While keeping the actual (saved) lora weight value large, always weaken the weight by a certain proportion during training to make the weight value appear smaller. This "weakening ratio" is Network Alpha. The smaller the Network Alpha value, the larger the weight value of the stored LoRA neural networks. Don't adjust the default parameters at will in professional mode, which may lead to even more outrageous results. If you are not sure what a certain parameter is for, try not to adjust it. Newbie recommends using basic mode. Scrambling labels: Usually, the earlier the word in the title, the more important it is. Therefore,if the order of the words is fixed, the following words may not be well learned, or the previous words may have unexpected associations with image generation. By randomly changing the order of the words each time the image is loaded, this bias can be corrected. Keep the Nth token: The first n words specified will always remain at the front of the title, which can be used to set the trigger word. Here, "word" refers to text separated by commas. Regardless of how many words the separated text contains, it is considered as "1 word". For example, for "black cat, eating, sitting", "black cat" is considered as 1 word. Noise offset: add noise offset in training to improve the generation of very dark or very bright images, not too large, try to be below 0.2 Multi-resolution noise attenuation rate: Multiple resolution noise iterations: Convolution layer dimension: Convolution layer Alpha value: Prompt word, sampling algorithm: The prompt word and sampling algorithm here are preset images for each version saved by Epoch, used to preview the training effect of the model 3.Training processBecause a machine can only run one Model Training task at the sanhe time, please be patient when facing possible queuing situations. We will prepare the training machine foyou assoon as possible. You can also perform staggered training at night 4.Model testingCurrently, after finding a suitable model for the example diagram and publishing it, do not upload display images (no images for display will be distributed to the homepage).After the deployment is completed, you can test your own model on the workbench. 5.Model release/download/retrain After the training is completed, you will see four preview images ofeach Epoch. You can choose the satisfactory works to publish in Tensorart or save them locally. If you are not satisfied with this training, you can view the training parameters and retrain in the upper right cornher. The specific method of adjusting parameters can be found in the above instructions. 

Training an AI model may sound daunting, but it’s surprisingly straightforward when broken down into clear steps. Whether you're a beginner or looking to refine your skills, this guide walks you through the process from creating datasets to fine-tuning settings. THIS IS THE MODEL PAGE : https://tensor.art/models/806678236659647115/CHRISTMAS-UGLY-SWEATER-PATTERN-V9Step 1: Build Your DatasetA quality dataset is the backbone of any successful AI model. Here’s how you can create one:Source Images: Gather images from reliable sources like Pinterest, stock image websites, your personal photo gallery, or even AI-generated outputs. Ensure you have permission to use the images, especially for commercial purposes.Focus on Quality:Use clear, sharp images.Avoid images with noise, blur, or watermarks.Size doesn’t have to be massive, but clarity is key.Example: For this guide, let’s say you’re building a dataset of seamless patterns inspired by ugly sweaters. Carefully curate high-quality images that fit this niche.Step 2: Caption Your DatasetGood captions make a significant difference in training outcomes. A well-captioned dataset ensures your model understands the nuances of your images.Tips for Effective Captioning:Write captions manually for precision.Use automated captioning tools sparingly and always review their output.Be descriptive but concise, capturing key details like color, style, or patterns.Example Caption:For an image of a red-and-green holiday sweater with reindeer motifs, your caption might read:“Seamless pattern of a red-and-green knitted sweater with reindeer and snowflake designs.”Manually crafting captions might take more time, but the payoff is better accuracy in your model's outputs.Step 3: Set Parameters and Configure TrainingOnce your dataset is ready, it’s time to train your model. Using platforms like Tensor.art simplifies this process.For Beginners:Start with default settings. These are optimized for general use and save you the hassle of configuring every parameter manually.For Advanced Users:Experiment with parameters such as learning rate, batch size, and epoch count to refine your model.Bonus TipsTest Regularly: As your model trains, run tests to ensure it’s learning correctly. This helps identify issues early.Iterate: Training is an iterative process. Don’t hesitate to tweak and retrain if the results aren’t up to par.Document Your Process: Keep notes on what works and what doesn’t. This saves time in future projects.Final ThoughtsTraining an AI model involves careful preparation and a bit of patience, but the results are worth the effort. By curating a high-quality dataset, writing thoughtful captions, and fine-tuning settings, you’ll be on your way to creating a model that performs exactly as you envision.Dive in, experiment, and watch your AI-powered creativity take flight!

Basics of creation LoRA. Model Training. Everything has a dark side.What does model training look like on the TensorArt platform?Why exactly you might need your own LoRA model?I'll briefly go over how to create your own LoRA model?What is the dark side of Model Training, and what does the possible blocking of the Telegram messenger in South Korea have to do with it?1. What does model training look like on the TensorArt platform?The training interface on the site looks like this.2. Why exactly you might need your own LoRA model?You need your LoRA model to replicate someone else's drawing style. Teach the model to draw certain objects, items that you need. And get similar images in the quantity you need. The use of models is limited only by your imagination. You can teach the model to draw anything, and then with its help improve the quality of your images. The accessibility and ease of learning provided by "TensorArt" makes "Model Training" accessible to everyone. Even to an inexperienced user. The model can also be taught to make some deepfakes of real people, images that look quite realistic. But this is the dark side "Model Training" and I will talk about it at the end of the article.3. How to create your own LoRA model?First, select the basic AI model, this is the foundation on which the model will be built (from there, the necessary missing elements and training algorithms).Once you have chosen the base model. Upload the images on which the training will take place (images that the model will learn to copy, style, drawing, specific objects, and the like). Try to find images with good resolution. Look for images with a resolution of more than 1000 pixels. :))For free users, training is available on no more than a hundred images, but do not be fooled.Training even on 15 images is quite expensive. As an example, training on 15 images with Repeat = 25 and Epoch = 10 on Pony Diffusion cost me 337 energy. "Total Repetitions Per Image: 250; Total Steps: 3750" Yes, Repeat = 25, this is the maximum for regular users. I think that 100 images and Repeat = 25 for training is more than enough for a regular user. Yes, PRO users can train the model on 1000 images and without a limit on Repeat. These are truly professional parameters, not for amateurs.After you upload the images, the AI ​​will assign tags to them (what it saw in your images). I advise you to carefully check the tags (keywords) on each image. Remove inappropriate keywords and add as many of your own as possible. Be careful when choosing words, do not write what is not in the picture.After carefully checking the tags, select the Repeat parameters (affects the quality of training, how many times it draws the same thing during training).Select the Epoch parameter. These are some evolutionary leaps in training. You can imagine it like this: Epoch one is the parent, and Epoch two is their child, and so on for generations. With this parameter, you select the number of generations.Fill in the "Trigger words" parameter. These are the words that launch your model in the prompt. An example of such a word: Old_postcards. If such a word is present, the model will launch and begin working, even if the value on the LoRA slider is small. Come up with words that characterize your model, no more than three.Next, fill in the "Model Effect Preview Prompt" parameter. Here, enter the maximum number of tags (keywords) that are on your images. During the check, you have already added them and deleted unnecessary ones. Now put them together and enter them here.Next, click the start training button and wait a while. When the training is over, select the most successful generation (you will have 4 pictures representing the Epoch), usually this is the last one. Click the publish button next to the Epoch.Next, fill out the required form to publish the model, everything the model is ready.In this article, I will not describe the intricacies of filling out the model publication form and more professional specific parameters for training. Since I promised it would be brief, but quite a lot has already come out. If you are interested, I will make a separate article on this topic. Write a feedback in the comments about it.4. What is the dark side of Model Training?Also, the model can be trained to make some deepfakes of real people, images that look quite realistic. And here we go to the dark side of Model Training. So recently, many articles have been published about the “Deepfake Epidemic”.In short, these articles were about journalists finding many Telegram channels where users could order deepfakes created from photos of real women via a Telegram bot. The first two deepfakes were offered for free, and each subsequent one was offered for money. The channels allowed users to upload an image of any woman. Then they found other Telegram channels collecting deepfakes of women of different professions - athletes, teachers, medical workers, and military personnel.As a result of these articles, South Korea seriously considered blocking the messenger completely. Telegram is not the only platform involved in this, as the attackers used X: to access a closed Telegram channel, and photos for deepfakes were often taken from Instagram.In 2020, the creator of one of the networks of such Telegram channels from South Korea was sentenced to 40 years in prison.The conclusion is, do not engage in such Deepfakes 18+, and also be careful when uploading photos to Instagram. They can make a Deepfake 18+ based on you.Every technology has its dark side, but this is not a reason for thoughtless complete bans, as well as hasty actions. It is necessary to develop regulations so that they affect only criminals. And will not affect honest users and developers. Only a person decides whether a tool in his hands will be an instrument of crime.

Greetings, my fellow Tensorian artists! 😊🤗😁@sfsinspiredwritr here! My Tensorian friend @digital_daydreamer had asked for my advice on training for models, and I gave him/her a helping hand! It's so detailed and invaluable, that I decided to give ALL Tensorians here these 10 personal tips on how to make models online with Tensor Art! (Note: These are the things that I have figured SO FAR in my Tensor journey up 'til now, and it's all from scratch!)Enjoy! 😄Tip #1: In my experience of training models (online on Tensor Art), a good margin for how many images to use (especially, for the model to work nicely when it's ran) is like 15-22 images. Any fewer or greater number will overall lessen the model's quality.Tip #2: For the images, Tensor Art supports images in the formats: .png, .jpg, .jpeg, and .webp (Chrome HTML document). I find these images by searching for the model topic on Google (for example, "digimon tai x sora") and then heading to the "Images" tab to see only photos of the topic. I click on an image I want to use, and then right-click on it to "Save image as". This is how I download the image to my computer.Now, ALWAYS check for the image format after you download, from whatever your file manager is called on your device. .avif, .jpg_large, nor .jfif formatted images are NOT supported by Tensor Art, so make sure to keep those out of your image count!Tip #3:Pony is THE #1 type of Model to use as a Base to train your models on online! The image quality is unmatched by any base model type, IMO! Good base models to use incl. Nova Anime XL (v.5.0 or v.6.0, haven't tried the newer versions yet) and Pony - Disney Pixar Style (V2).Tip #4:The steps, how many times the model repeats an image for training, is IMPORTANT to set up! Generally you keep the steps in the range of 24-27 steps for the model to train the images on.Tip #5:You should keep the image prompt as 'loosely detailed" as possible; my image prompt in online training would look something like this:(Trigger words) - digimon_adventure, mimi_tachikawa, (1girl, 1_solo) :(Image standard terms) - (RAW Photo, best_quality, high_definition, high_res, masterpiece, veryGOODImage_positive) (Beautiful, extremely detailed)_image, (BEST, amazing, incredible, highly detailed, anime-style)_quality, (Clear & not blurry, beautiful)_image(Optional details) - simple_background, (upper_body(_shot))(Main Prompt) There is a teenage girl resembling Mimi Tachikawa. She has long and light brown hair, (amazing, anime_style, well-drawn) brown eyes, and fair skin. She is looking at the viewer with a smile.This is how my training prompt looks! Simple enough for the AI to train the images! 🙂Tip #6:The steps are 24-27, and the epoch is just 1, a set of 4 images.This is the cheapest way I practice to make images, and with what I spoke on the steps and the image prompt, the model will become fantastic to use using this ratio!*The more images you want to use, the lower the steps should be for conservative generation.Tip #7:The model's initial images, when its training is done, will look ... UNPROFESSIONAL for the most part.Don't be discouraged! Consider these first 4 images the "rough draft", and just Remember This: The model's first images are not the goal to completing the model, it's the images that you run AFTER the model is published!Essentially, I just replace the first images with 10-12 of NEW images to show off for the model! That's the "presentation" secret (Shh 🤫)!Tip #8:Upon running the newly published model, this is the part where you come up with: more details for the prompt, the negative prompt, the strength ('detail weight') of the model, the CFG scale, and whatever else you can mess with to make good images with the model!Once you have this figured out, then you can "Edit Model" & place your "base" prompt & negative prompt & strength & CFG scale & other details into the description of the model itself so that the users running your model (and yourself 😉 wink-wink) can have some more ease and AMAZING use with it!Tip# 9:The models will have to go into projects to be showcased to the Tensor community, of course!To make a project, you go to "+ Host my model" on your Profile page, then go to "Create a project" to go to a page that does just that! You make its name, type (mines are usually Image & LoRA), channel (Anime, Illustration, etc.), tags (e.g. "SORA TAKENOUCHI", "DIGIMON ADVENTURE", etc.), description (where I recommend some of my tips for users to use the model well), resource (usually "Original"), view scope (usu. "Public"), and permissions you give to users (I allow everything except those under "Commercial use"; I don't care about those), and then you "Create"!Tip #10:Instead of using the model you used to train the images to run new images to present, you can use a DIFFERENT ONE to make the new images of the model!For example, when I train models for the DIGIMON ADVENTURE characters' Defined versions, the base model for training is Nova Anime, but for running images it's Disney Pixar Style! For their anime versions, it's the reverse! In my experience, using a different model for running aside from the one used for training really brings out the detail and quality of the images, better than using the same one! (Nothing against using the same one, though; this would still work!)And there you have it! I had a lot to say, and (again) I had to learn these priceless pieces of knowledge by MYSELF, and thankfully you new and seasoned Tensor artists should have more ease in training your LORA models online here on Tensor Art 🥰!I hope these tips are of great help to you (let me know if you're confused), and let us all grow into amazing Tensor Artists together!P.S. When I learn something else new, I'll be sure to post "mini-articles" as updates for my sharing these experiences with you all, Tensor Art community!Until another time, Tensorians! Many good wishes to you all, fellow artists and creators! 🤗🤩🥰😍😊😁😀

IntroductionIn the rapidly evolving field of AI image generation, techniques such as LoRA (Low-Rank Adaptation) and LoKr have emerged as powerful methods for fine-tuning large models like Stable Diffusion. Understanding the differences between these methods, their advantages and disadvantages, and how they can be applied effectively is crucial for practitioners aiming to generate high-quality images efficiently.In this article, we will delve into the distinctions between LoRA and LoKr, explore the strengths and weaknesses of each approach, and provide a detailed explanation of LoKr (which is also known as LyCORIS). We will focus on how LoKr offers significant advantages in AI image generation.Understanding LoRAWhat is LoRA?LoRA, or Low-Rank Adaptation, is a technique designed to efficiently fine-tune large pre-trained models by injecting trainable low-rank matrices into their architecture. Instead of updating all of the parameters of a model during fine-tuning, LoRA introduces additional low-rank weight matrices that capture task-specific information. This approach significantly reduces the number of parameters that need to be updated, leading to lower computational costs and memory requirements.Advantages of LoRAEfficiency: LoRA reduces the computational resources required for fine-tuning by updating only a small number of parameters.Memory Footprint: The additional low-rank matrices consume less memory compared to full fine-tuning.Speed: Faster training times due to fewer parameters being optimized.Disadvantages of LoRALimited Expressiveness: The low-rank matrices may not capture complex patterns as effectively as full fine-tuning.Performance Trade-offs: In some cases, LoRA may result in slightly lower performance compared to methods that fine-tune all parameters.Understanding LoKr (LyCORIS)What is LoKr?LoKr, standing for Low-Rank Kronecker product adaptation, is an advanced fine-tuning technique that extends the principles of LoRA by incorporating Kronecker products into the adaptation process. LoKr is part of the LyCORIS framework (Low-Rank Compression via Rank-One updates and shared Subspace), which is designed to improve the efficiency and effectiveness of model adaptation in AI image generation tasks.LoKr introduces more expressive adaptation layers by utilizing Kronecker products, allowing the model to capture more complex interactions and patterns within the data without significantly increasing the number of parameters.Advantages of LoKrEnhanced Expressiveness: By using Kronecker products, LoKr can model more complex relationships in the data.Parameter Efficiency: Achieves higher performance without a proportional increase in parameters compared to full fine-tuning.Improved Image Quality: Particularly effective in capturing detailed textures and styles in AI-generated images.Disadvantages of LoKrComplexity: The implementation of Kronecker products adds complexity to the adaptation process.Computational Overhead: Slightly higher computational requirements than LoRA due to the more complex operations.Differences Between LoRA and LoKrAdaptation Methodology:LoRA uses low-rank matrices added to the model's weights to capture task-specific information.LoKr extends this by incorporating Kronecker products, allowing for modeling higher-order interactions.Expressiveness:LoRA may struggle with capturing complex patterns due to the limitations of low-rank representations.LoKr provides enhanced expressiveness, enabling the model to learn more intricate patterns.Parameter Efficiency:LoRA is highly parameter-efficient but may sacrifice some performance.LoKr balances parameter efficiency with improved performance, offering better results without a significant increase in parameters.Computational Requirements:LoRA requires less computation and is faster to train.LoKr has slightly higher computational demands but offers better performance for complex tasks.The Advantages of LoKr (LyCORIS) in AI Image Generation1. Superior Detail CaptureLoKr excels in capturing fine-grained details in images. By leveraging Kronecker products, it can model complex spatial patterns and textures that are often present in high-quality images. This leads to more realistic and detailed image generation.2. Improved Style TransferIn tasks involving style transfer or adaptation to new artistic styles, LoKr's enhanced expressiveness allows it to better capture the nuances of different styles. This results in generated images that more faithfully represent the desired aesthetic.3. Efficient AdaptationLoKr achieves a balance between parameter efficiency and performance. It allows for fine-tuning models to new tasks without the need to update all parameters, saving computational resources while still delivering high-quality results.4. FlexibilityThe approach can be applied to various layers within the model, providing flexibility in how and where the adaptation occurs. This allows practitioners to tailor the fine-tuning process to the specific needs of their task.Detailed Insights into LoKr (LyCORIS)While we won't cover installation or practical steps, understanding how LoKr works at a deeper level can help practitioners make informed decisions about its use.Kronecker Products in LoKrThe Kronecker product is a mathematical operation that produces a block matrix from two smaller matrices. In the context of LoKr, it allows for the creation of adaptation matrices that are capable of modeling higher-order interactions without a significant increase in parameters.By utilizing Kronecker products, LoKr can inject more expressive transformations into the model's layers. This enables the model to learn complex relationships within the data, which is particularly beneficial for image-generation tasks that require capturing intricate patterns and textures.Parameter Efficiency and PerformanceLoKr maintains a balance between the number of parameters and the performance of the model. By carefully designing the adaptation matrices using Kronecker products, it achieves improved expressiveness without the need for a large number of additional parameters.This efficiency is crucial in scenarios where computational resources are limited but high performance is still required.Applicability to Stable DiffusionLoKr is especially suitable for fine-tuning Stable Diffusion models. It enhances the model's ability to generate high-quality images by effectively adapting to new styles and subjects. The flexibility of LoKr allows it to be integrated into various parts of the model, providing a powerful tool for practitioners in the AI image generation field.ConclusionIn the field of AI image generation, both LoRA and LoKr offer valuable methods for fine-tuning large models efficiently. While LoRA provides a simple and resource-efficient approach, LoKr (LyCORIS) extends these capabilities by introducing Kronecker products to capture more complex patterns and interactions.LoKr stands out for its ability to enhance image quality, capture detailed textures, and adapt to new styles with greater fidelity. Its advantages make it a compelling choice for practitioners seeking to push the boundaries of AI-generated imagery.By understanding the differences between LoRA and LoKr, and appreciating the strengths of LoKr in AI image generation, practitioners can make informed decisions about which technique best suits their needs.

A step-by-step guide on how to train a LoRA; part 2!Warning: This guide is based on Kohya_SSThis guide REQUIRES that you read "How I LoRA: A beginners guide to LoRA training | Part 1: Dataset Prep."This guide CAN be ported to Tensor.art's trainer; if you know what you are doing.This guide is an (almost) 1:1 of the following guide: https://civitai.com/articles/3522/valstrixs-crash-course-guide-to-lora-and-lycoris-trainingEdits were made to keep it short and only dive into the crucial details. It also removes a lot of recommendations I DO NOT follow.; for more advanced information, please support the original guide. If you want to do things MY way, keep reading.THE SETTINGS USED ARE BASED ON SDXL, DO NOT FOLLOW IF YOU ARE TRAINING ON V-PRED OR 1.5Training: BasicsNow that you have your dataset, you need to actually train it, which requires a training script. The most commonly used script, which I also use, are the Kohya Scripts. I personally use the Kohya-SS GUI, a fork of the SD-Scripts command line trainer.Once you have it installed and open, make sure you navigate to the LoRA tab at the top (it defaults to dreambooth, an older method.)There are a lot of things that can be tweaked and changed in Kohya, so we'll take it slow. Assume that anything I don't mention here can be left alone.We'll go down vertically, tab by tab.Accelerate Launch This tab is where your multi-gpu settings are, if you have them. Otherwise, skip this tab entirely, as the defaults are perfectly fine. Training precision is also here, and should match your Save precision in the following tab, but you won't touch it otherwise.ModelThis tab, as you've likely guessed, is where you set your model for training, select your dataset, etc.Pretrained model name or path:Input the full file path to the model you'll use to train.Trained Model output name:Will be the name of your output file. Name it however you like.Image folder (containing training images subfolders):Should be the full file path to your training folder, but not the one with the X_. You should set the path to the folder that folder is inside of. Ex: "C:/Training Folders/Concept Folder/".Underneath that, there are 3 checkboxes:v2: Check if you're using a SD 2.X model.v_parameterization: Check if your model supports V-Prediction (VPred).SDXL Model: Check if you're using some form of SDXL, obviously.Save trained model as:Can stay as "safetensors". "ckpt" is an older, less secure format. Unless you're purposefully using an ancient pre-safetensor version of something, ckpt should never be used.Save precision:"fp16" has higher precision data, but internally has smaller max values. "bf16" holds less precise data, but can use larger values, and seems faster to train on non-consumer cards (if you happen to have one). Choose based on your needs, but I stick with fp16 as the higher precision is generally better for more complex designs. "float" saves your LoRA in fp32 format, which gives it an overkill file size. Niche usage.MetadataA section for meta information. This is entirely optional, but could help people figure out how to use the LoRA (or who made it) if they find it off-site. I recommend putting your username in the author slot, at least.FoldersAs simple as it gets: Set your output/reg folders here, and logging directory if you want to.Output folder:Where your models will end up when they are saved during/after training. Set this to wherever you like.Regularization directory:Should be left empty unless you plan to use a Prior Preservation dataset from section 3.5, following a similar path to the image folder. Ex: "C:/Training Folders/Regularization/RegConceptA/".ParametersThe bread-and-butter of training. Mostly everything we'll set is in this section: Don't bother with the presets, most of the time.Lora Type: StandardTrain Batch Size:How many images will be trained simultaneously. The larger this number the more VRAM you will use, don't go over 1 if you have low VRAM.Max train steps:RECOMMENDED. Forces training to stop at the exact step count provided, overriding epochs. Useful if you want to stop at a flat 2000 steps or similar. 3000 steps is my recommended cap.Save every n epochs:RECOMMENDED. Saves a LoRA before it finishes every X number of epochs you set. This can be useful to go back to and see where your sweet spot might be. I usually keep this at 1, saving every epoch.Your final epoch will always be saved, so setting this to an odd number can prove useful, such as saving every 3 epochs with a 10 epoch training will give you epochs 3, 6, 9, & 10, giving you a fallback right at the end if it started to overbake.Cache latents & Cache latents to disk:These affect where your data is loaded during training. If you have a recent graphics card, "cache latents" is the better and faster choice which keeps your data loaded on the card while it trains. If you're lacking VRAM, the "to disk" version is slower but doesn't eat your VRAM to do so.Caching to disk, however, prevents the need to re-cache the data if you run it multiple times, so long as there wasn't any changes to it. Useful for tweaking trainer settings. (Recommended).The next settings are calibrated for low VRAM usage; read the original guide if you got VRAM to spare. Anything highlighted was changed for maximum VRAM optimization.LR Scheduler: Cosine With RestartsOptimizer: AdamW8bit.Optimizer extra arguments: weight_decay=0.01 betas=0.9,0.99"Learning Rate": 0.0001As a general note, the specific "text encoder" and "unet" learning rate boxes lower down will override the main box, if values are set in them.LR warmup (% of total steps): 20"LR # cycles": 3"Max resolution": 1024,1024Enable buckets: TrueMin/Max bucket resolution: 256; 2048"Text Encoder & Unet learning rate": 0.001; 0.003No half VAE: Should always be True, imo, just to save you the headache.Network Rank & Network Alpha: 8 / 8Your Alpha should be kept to the same number as your Rank, in most scenarios.Network Dropout:Recommended, but optional. A value of 0.1 is a good, universal value. Helps with overfitting in most scenarios.Advanced (Subtab)We won't touch much here, as most values have niche purposes.Gradient accumulate steps: 1Prior loss weight: 1Keep n tokens:For use with caption shuffling, to prevent the first X number of tags from being shuffled.If using shuffling, this should always be 1 at minimum, which will prevent your instance token from being thrown around.Clip skip:Should be set to the clip skip value of your model. Most anime & SDXL models use 2, most others use 1. If you're unsure, most civit models note the used value on their page.Full bf16 training: FalseGradient Checkpointing: TrueShuffle Caption: TruePersistent Data Loader: FalseMemory Efficient Attention:Use only if you're not on a Nvidia card, like AMD. This is to replace xformers CrossAttention.CrossAttention:xformers, always. (As long as you're on a Nvidia card, which you really should be.)If for whatever reason you can't use xformers, SDPA is your next best option. It eats more ram and is a bit slower, but it's better than nothing.Color augmentation:Do not.Flip Augmentation: FalseMin SNR Gamma: 1Debiased Estimation Loss: FalseBucket resolution steps: 64Random crop instead of center crop: FalseNoise offset type: MultiresMultires noise iterations: 6Multires noise discount: 0.3IP noise gamma: 0.1And that's everything! Scroll to the top, open the "configuration" dropdown, and save your settings with whatever name you'd like. Once you've done that, hit "start training" at the bottom and wait! Depending on your card, settings, and image count, this can take quite some time.Here are some visuals:Once training begins:THIS IS IT FOR PART 2: Training Basics!

介绍在快速发展的 AI 图像生成领域，LoRA（低秩适应）和 LoKr 等技术已经成为微调大型模型（如 Stable Diffusion）的强大方法。了解这些方法之间的区别、它们的优缺点以及如何有效应用，对于希望高效生成高质量图像的从业者来说至关重要。本文将深入探讨 LoRA 和 LoKr 之间的区别，分析每种方法的优缺点，并详细解释 LoKr（也称为 LyCORIS）。我们将重点关注 LoKr 在 AI 图像生成中的显著优势。了解 LoRA什么是 LoRA？LoRA，即低秩适应（Low-Rank Adaptation），是一种旨在高效微调大型预训练模型的技术。它通过在模型架构中注入可训练的低秩矩阵，而不是更新模型的所有参数。在微调过程中，LoRA 引入额外的低秩权重矩阵，以捕获特定任务的信息。这种方法大大减少了需要更新的参数数量，从而降低了计算成本和内存需求。LoRA 的优势高效性：LoRA 仅更新少量参数，减少了微调所需的计算资源。内存占用小：额外的低秩矩阵相比全面微调消耗更少的内存。训练速度快：由于优化的参数较少，训练时间更短。LoRA 的劣势表达能力有限：低秩矩阵可能无法有效捕获复杂的模式。性能权衡：在某些情况下，LoRA 的性能可能略低于全面微调所有参数的方法。了解 LoKr（LyCORIS）什么是 LoKr？LoKr，即低秩克罗内克积适应（Low-Rank Kronecker product adaptation），是一种先进的微调技术，通过在适应过程中引入克罗内克积来扩展 LoRA 的原理。LoKr 是 LyCORIS 框架（通过秩一更新和共享子空间实现的低秩压缩）的一部分，旨在提高 AI 图像生成任务中模型适应的效率和效果。LoKr 通过利用克罗内克积引入更具表现力的适应层，使模型能够在不显著增加参数数量的情况下，捕获数据中更复杂的交互和模式。LoKr 的优势增强的表达能力：通过使用克罗内克积，LoKr 能够建模数据中更复杂的关系。参数效率：相比全面微调，在不成比例增加参数的情况下实现更高的性能。改进的图像质量：在捕获 AI 生成图像的细节纹理和风格方面特别有效。LoKr 的劣势复杂性：克罗内克积的实现增加了适应过程的复杂性。计算开销：由于更复杂的操作，计算需求略高于 LoRA。LoRA 和 LoKr 的区别适应方法LoRA：使用添加到模型权重中的低秩矩阵来捕获特定任务的信息。LoKr：通过引入克罗内克积，能够建模高阶交互。表达能力LoRA：由于低秩表示的限制，可能难以捕获复杂的模式。LoKr：提供了增强的表达能力，使模型能够学习更复杂的模式。参数效率LoRA：高度参数高效，但可能牺牲一些性能。LoKr：在参数效率和性能之间取得平衡，提供更好的结果，而不显著增加参数。计算需求LoRA：需要较少的计算，训练速度更快。LoKr：计算需求略高，但在复杂任务中提供更好的性能。LoKr（LyCORIS）在 AI 图像生成中的优势1. 优异的细节捕捉LoKr 在捕捉图像的细粒度细节方面表现出色。通过利用克罗内克积，它可以建模图像中复杂的空间模式和纹理，生成更逼真和详细的图像。2. 改进的风格迁移在涉及风格迁移或适应新艺术风格的任务中，LoKr 的增强表达能力使其能够更好地捕捉不同风格的细微差别，生成的图像更忠实地呈现所需的美学效果。3. 高效的适应性LoKr 在参数效率和性能之间取得平衡。它允许在不更新所有参数的情况下，将模型微调到新任务，节省计算资源，同时仍然提供高质量的结果。4. 灵活性该方法可以应用于模型中的各种层，提供了适应发生位置和方式的灵活性，使从业者能够根据任务的具体需求定制微调过程。LoKr（LyCORIS）的详细解析虽然我们不涉及安装或实际操作步骤，但深入了解 LoKr 的工作原理可以帮助从业者做出明智的决策。LoKr 中的克罗内克积克罗内克积是一种数学运算，可从两个较小的矩阵生成一个块矩阵。在 LoKr 的背景下，它允许创建能够建模高阶交互的适应矩阵，而不会显著增加参数数量。通过利用克罗内克积，LoKr 可以在模型的层中注入更具表现力的变换，使模型能够学习数据中的复杂关系。这对于需要捕捉复杂模式和纹理的图像生成任务特别有益。参数效率和性能LoKr 在参数数量和模型性能之间保持平衡。通过使用克罗内克积精心设计适应矩阵，它在无需大量额外参数的情况下，实现了增强的表达能力。在计算资源有限但仍需要高性能的情况下，这种效率至关重要。适用于 Stable DiffusionLoKr 特别适合微调 Stable Diffusion 模型。它通过有效地适应新风格和主题，增强了模型生成高质量图像的能力。LoKr 的灵活性允许其集成到模型的各个部分，为 AI 图像生成领域的从业者提供了强大的工具。结论在 AI 图像生成领域，LoRA 和 LoKr 都提供了高效微调大型模型的有价值方法。LoRA 提供了一种简单且资源高效的方法，而 LoKr（LyCORIS）通过引入克罗内克积来捕捉更复杂的模式和交互，扩展了这些能力。LoKr 在增强图像质量、捕捉细节纹理以及更高保真度地适应新风格方面表现突出。其优势使其成为希望推动 AI 生成图像边界的从业者的理想选择。通过了解 LoRA 和 LoKr 之间的区别，并认识到 LoKr 在 AI 图像生成中的优势，从业者可以根据自身需求做出最佳选择。

Attention to detail in the prompt is important to produce beautiful images related to Christmas.We've put together a list of words to help you get into the Christmas spirit and create dreamy and beautiful images.By using these in combination, you are sure to be able to create a magical and warm Christmas scene. 　1. Main theme　- Christmas tree　- Santa Claus　- Snowy landscape　- Christmas lights　- Cozy fireplace　- Gift-wrapped presents　- Snow-covered village　- Reindeer　- Gingerbread house　2. Decorations and accessories　- Ornaments　- Garland　- Wreath　- Baubles　- Holly and ivy　- Christmas stockings　- Snowflakes　- Candy canes　- Bells　3. Lighting effects　- Warm glowing lights　- Sparkling fairy lights　- Candlelit ambiance　- Radiant glow　- Soft bokeh effect　- Starry night　- Lantern-lit　4. Color effects　- Warm hues　- Rich reds　- Forest greens　- Snowy whites　- Golden tones　- Frosty blues　- Silver sparkle　5. Characters and animals　- Santa’s elves　- Festive animals　- Polar bear with scarf　- Penguins in winter hats　- Adorable snowman　6. Scenery and places　- Cozy cabin　- Snowy forest　- Mountain village　- Christmas market　- ​​Enchanted winter scene　- Icicles on trees　- Ice rink with skaters　7. Texture and atmosphere　- Velvet textures　- Frosted glass effect　- Glittering snow　- Cozy and warm atmosphere　- Nostalgic charm　- Festive elegance　- Whimsical charm　8. Other keywords　- Magical holiday　- Winter wonderland　- Silent night　- Joyful celebrations　-Family gathering　- Childlike wonder　- Frost-kissed scenerytogether with a combination of these words"Beautiful", "Wonderful", "Warm", "Festive" etc.Adding keywords to the prompt will help you generate images full of Christmas charm.

What is LoRA?Low-Rank Adaptation (LoRA) is a technique that allows efficient fine-tuning of AI models without the need to retrain the entire model. With LoRA, AI creators can adapt models in a resource-efficient way while maintaining high quality.Using Tensor Art AI, you can easily create LoRA models without deep knowledge of programming or machine learning. This guide will cover all the steps involved in creating LoRA in Tensor Art AI, from data preparation to implementation in your creative workflow.Why Use LoRA in Tensor Art AI?Tensor Art AI offers many benefits for AI creators who want to build LoRA models:✅ Resource Efficient – No need for a high-spec computer.✅ Easy to Use – User-friendly interface simplifies training.✅ Flexible and Customizable – Adapt AI models to your needs.✅ Supports Multiple Models – Can be used with Stable Diffusion and other AI models.With these advantages, Tensor Art AI is an ideal choice for creating AI with unique characteristics.Preparation Before Creating LoRA1. System and Device RequirementsA Tensor Art AI account (register if you don’t have one).A stable internet connection.A compatible browser (Google Chrome recommended).2. Collecting High-Quality DatasetsUse high-resolution images (at least 512x512 pixels).Vary styles and expressions for better results.Avoid images with watermarks or distracting elements.Remove duplicate or irrelevant images from the dataset.Save images in JPEG or PNG format and organize them systematically.Steps to Create LoRA in Tensor Art AI1. Access Tensor Art AI DashboardOpen the Tensor Art AI website and log in to your account.Select Training LoRA from the Create tab, then choose Online Training.2. Upload Your DatasetClick the Upload Dataset button.Select and upload your prepared images.Label and categorize the images for better pattern recognition.Choose the model to use (Flux, Stable Diffusion 1.5, SDXL, etc.).3. Configure Training ParametersLearning Rate: 1e-4 (recommended for beginners, can be adjusted).Batch Size: 4-8 (depending on available GPU capacity).Epochs: 10-20 (for optimal results).Resolution: 512x512 or higher if needed.4. Start Training ProcessClick Start Training and wait for the process to complete.This process may take several hours, depending on the dataset and system specifications.Tensor Art AI will display training progress such as loss function and model accuracy.5. Publish Your ModelOnce training is complete, publish your LoRA model via Share directly or Host my Model.Your LoRA model is now ready for use on the Tensor Art AI platform.Using LoRA Models Locally1. Downloading Your LoRA ModelGo to Training History in the dashboard.Select the trained model.Click Download LoRA.2. Using LoRA for Image GenerationOn Tensor Art AI: Go to Generate Image, select a base model (Stable Diffusion, etc.), and add LoRA.On Third-Party Applications: LoRA can be used in software such as Automatic1111 or ComfyUI.Experiment with Prompts: Use different prompt combinations to get unique and varied results.Tips for Optimal LoRA Results✔ Use Larger Datasets – More quality data results in a more accurate model. ✔ Experiment with Training Parameters – Try different combinations for optimal output. ✔ Use Pretrained Models as a Base – Speeds up the process and improves results. ✔ Join the AI Community – Engage in AI forums or the Tensor Art AI Discord for insights. ✔ Regularly Update Your Model – Add new datasets to improve image quality.LoRA in Tensor Art AI: Pros and Cons✅ ProsResource Efficient – No need for a high-spec computer.User-Friendly – Easy to train, even for beginners.Fast – Training is relatively quick, especially with CUDA-supported GPUs.Flexible – Works with various AI models.High Customization – More control over training parameters.Active Community – Support and additional insights available.❌ ConsDataset Quality Matters – Poor datasets can lower model performance.Internet Dependent – Upload and training require a stable connection.Training Time – While fast, training still takes time, especially for large datasets.Costs – Free credits are limited; optimal results require more credits.Possible Errors – Some technical issues may arise during training.ConclusionCreating a LoRA model in Tensor Art AI is an exciting and easy process, even for beginners. By following these steps, you can develop AI models tailored to your creative needs.Feel free to experiment with different datasets and training parameters to achieve the best results. With consistent practice, you can create high-quality AI models for various creative projects.Happy experimenting and creating with LoRA in Tensor Art AI!HIJAB FUSION

To create an anime LoRA (Low-Rank Adaptation) model for Tensor Art, you’ll need to follow a process that involves dataset preparation, training, and fine-tuning. Here’s a step-by-step guide:1. Prepare Your DatasetImage Collection: Collect high-quality anime-style images. Make sure these images match the style you want to emulate (specific anime, studio, or theme).Image Resolution: Most LoRA models work best with images of uniform resolution, typically 512x512 or 768x768 pixels.Annotations/Labels: Depending on the platform you’re using, images may need annotations to help guide the AI in learning. Organize and label them clearly (e.g., specific character styles, poses, lighting).2. Preprocess ImagesImage Cleaning: Make sure the images are clean, with no watermarks or artifacts, and cropped to center the character or object of focus.Resolution Standardization: Resize images to the preferred resolution while maintaining aspect ratio.Augmentation (Optional): Apply augmentations like rotation, zoom, or color shifts to increase variety.3. Select a Base ModelChoose a suitable pre-trained base model for your LoRA training. This model can be an anime-style model like Stable Diffusion or a specific anime GPT-like architecture. Tensor Art may provide base models compatible with LoRA training.4. LoRA Training SetupLoRA Fine-Tuning: LoRA works by fine-tuning only part of a neural network's weights rather than retraining the entire model. Tools like Diffusers (Hugging Face) or Dreambooth can help set up LoRA fine-tuning with your dataset.Hyperparameters: Choose appropriate training settings (learning rate, batch size, steps) based on your hardware. LoRA models are less hardware-intensive than full model training, but you still need adequate GPU power.Learning Rate: A common learning rate is between 5e-5 and 1e-4 for LoRA models.Epochs: Set epochs between 2-5 to prevent overfitting.Batch Size: Keep the batch size balanced to avoid VRAM overflow (commonly between 4-8).5. Training the ModelRun Training: Start the fine-tuning process. Depending on the hardware and the size of your dataset, this could take a few hours to days.Monitor Loss: Watch the loss during training. If it plateaus or starts increasing, it may signal overfitting, and you may need to stop or adjust the training settings.6. Test & Fine-tuneOnce training is done, test your LoRA model by generating anime-style images using the model. Adjust any hyperparameters and retrain if the results aren’t satisfactory.Fine-tune until you achieve the desired style consistency.7. Export and ShareAfter training, export the LoRA model for usage in platforms like Tensor Art. Upload it as per the platform’s guidelines.Share or use your new LoRA model to generate anime-style images on Tensor Art.Tools and Platforms:Diffusers by Hugging Face: Supports LoRA training for text-to-image models.Kohya's LoRA Trainer: Popular in the AI community for LoRA training with anime datasets.Dreambooth: Can be used with LoRA techniques to fine-tune specific features in your model.Make sure to follow Tensor Art’s specific guidelines on model compatibility, as they may have their requirements for LoRA usage!

首先点击右上角的头像，在弹出的下拉框中选择我训练的模型，进入训练中心。如果之前有训练过模型，这里会看到许多训练任务。然后选择在线训练按钮进行一次训练。左侧是数据集窗口，默认没有任何数据。您可以上传一些图片作为数据集，或者上传一个数据集压缩包，压缩包可以包含标注文件，格式和kohya-ss一样，每个图片文件对应一个同名的标注文件txt。右边的模型主题中可以选择二次元人物、真实人物、2.5D、标准以及自定义。训练混元模型这里我们选择标准，在使用底模中选择混元1.2模型。混元模型使用了40depth的块，所以非常大，训练相对速度较慢，需要更高的学习率，默认使用4e-4，默认单张图片重复次数5，优化器AdamW。基础模式下参数选择，推荐单张图片重复次数5，轮数为16。上传一个处理好的数据集后，如果你的数据集标注中有人物名，可以不写触发词。否则你应该给你的模型起一个简单的触发词，例如人物名称或者风格名称。接着从数据集中选择一个标注文件作为预览提示词。如果你想使用专业模式，选择右上角按钮切换到专业模式。专业模式推荐学习率翻倍，然后使用cosine_with_restarts学习率调度器，优化器选择AdamW或者AdamW8bit。开启打乱标签(shuffle)，并且保持第1个token(如果你有一个人名触发词在第一个)关闭噪声偏移功能，卷积DIM和Alpha可以选择8和1。在样图设置中追加填写反向提示词，接下来就可以开始训练了。在训练队列中，你可以看到当前loss值变化表以及每轮epoch产生的4张样图。最后可以选择效果最好的epoch下载到本地或者直接在tensorart上发布。

Z-Image is a powerful and useful model, but the current Turbo version is not very friendly to training LoRa. Here are a few points to note, shared for those who need them:1. When training with TA, be sure to use the official version released by @通义万相, (Z-Image-Turbo), which allows De-distilled mode.2. Do not use Add Dataset Normalization, as it is not yet supported.3. Regarding the issue of the learning rate not increasing after the loss reaches 0.37, you can control the number of steps to avoid wasting computational resources. Around 5000-6000 steps should reach approximately 0.37; adjust according to your situation and training set.4. I mostly use the default settings for other configurations. Adjust them based on your experience.5. After LoRa is set up, test it with 20 steps and CGF: 1.5-2. Due to the distillation of the model, 9 steps and CGF: 1 may not have any effect.These are some of my experiences training the Z-Image Turbo model LoRa; I hope they are helpful.My Z-IMAGE LoRa link, feel free to test it.

Introduction(Character • Style • Multi-LoRA • Krea & Qwen notes)Hi friends! 🤗This is a friendly, experience-based walkthrough of how I build image LoRAs on Tensor.Art. I’m not a guru—just sharing what actually worked for me so you can skip a few potholes and have more fun. (This article is a merged compact version of previous three plus Flux Krea LoRA creation info~)We’ll cover:Character LoRA (Flux)Style LoRA (oil-painting vibe)Multi‑LoRA (several sub‑LoRAs inside one model)Flux Krea & Qwen LoRA (what’s the same, what’s different)Part 1. Character LoRA — “Skull Knight” (one of my early Flux projects) ⚔️💀I made a Flux character LoRA for a Skull Knight—not because the world needed it, but because I loved the vibe. It wasn’t my first-ever Flux LoRA, but it was one of the early ones and taught me a lot.ImagesTarget 15–40 solid images (I tried 12 for this case—doable, but more is safer).High quality + varied angles help a ton. Low-quality in → low-quality out (Flux is forgiving, but still).Core settings (Flux)Base: Flux.1 (default)Network: I like LoKr, but in project settings choose LyCORIS (LoKr selection has caused issues in some UIs; LyCORIS is the umbrella that supports LoKr/LoHA/LoCon, etc.).Trigger word: use a unique token (e.g., ek_sku11_kn1ght). I swap i/l with 1 to avoid collisions with normal words.Repeat / Epoch: e.g., 15 / 3 to save credits, then continue later if needed.Resolution: 1024×1024 is best for faces; 512×512 can still work for quick prototypes.Scheduler: cosine or cosine_with_restarts - 5% warming of total steps (constant/linear also fine).Optimizer: AdamW8bit (simple + memory‑friendly).Shuffle captions: ON. (OFF if # of image sets is low < 40)Keep N tokens: 1 (# of captions not to be shuffled).Noise offset: default 0.03 worked fine for me. (could be up to 0.05)conv_dim / conv_alpha (character): 4 / 1Labeling & promptsCaptioning: auto is fine; prepend the primary trigger to all captions (labeling tool helps).Add secondary descriptors at the end (e.g., “metal spikes”, “metallic surface”) to reinforce traits.Sample prompts during training should be simple so you can see the LoRA effect clearly.Training, picking epochs, publishingWatch epoch previews (4 images per epoch). Loss trends help, but your eyes win.If an epoch looks saturated (repeating) or artifacts creep in, stop early and save credits.Publish the best epoch. If you’re undecided, publish two versions (e.g., v1/v2 or pro/non‑pro) and let users pick.If you’re Pro, Continue Training from a good epoch is super handy.Result: even with 12 images at 512×512, I got a surprisingly usable LoRA. Flux does a lot of heavy lifting when your setup is sensible.Part 2. Style LoRA — Painting with Hopper 🎨This time I chased an Edward Hopper oil‑painting feel.Images30–100 images in a consistent style (I used 32 at 1024×1024).Add style‑specific caption hints like “flat colors”, “strong light–dark contrast”—these steer the vibe.Key differences from character LoRAStyle tends to learn fast → you often need fewer repeats/epochs.Example: repeat 10, epoch 3 was enough (pushing harder led to overfit or worse faces).conv_dim / conv_alpha (style): 8 / 2UNet LR: 0.0002 (default 0.0001 also okay; I nudged it for speed)Choosing the epochTreat it like tasting notes: Epoch 2 might be subtle and classy; Epoch 3 bold but riskier.A slightly higher-loss epoch can look better—trust visuals over numbers.Test strength: I usually try 0.8–1.0. That’s where most styles sing.Part 3. Multi‑LoRA — Several flavors in one 🍱Goal: train multiple sub‑LoRAs into a single “combo” model so you can call specific sub‑styles/characters by trigger.Folder idea:ComboChar/├── character_A/ (trigger: charA, 30 imgs)├── character_B/ (trigger: charB, 40 imgs)└── character_C/ (trigger: charC, 35 imgs)Caption pattern (at the start):ComboChar, charA or ComboChar, charB …Training switchesShuffle captions: TrueKeep N tokens: 2 (so those two triggers aren’t learned as content)After trainingUse ComboChar for the blended vibe, or ComboChar + charA to force a specific sub‑LoRA.During training, previewing all variants is limited; I check candidates after training, then re‑train if needed.I’ve used this trick for things like RPG Booster, Yellowstone/Yosemite, Winter Resort—packing variety into one model is surprisingly practical (and fun).Part 4. Flux Krea LoRA — Same recipe, new flavor 🍜Flux Krea popped up as the “new flavor” of Flux, so of course I had to try it with a Yor Forger LoRA.👉 The surprise? It’s basically Flux with a small twist.Network: LoRA only (no LoKr option here)network_dim / alpha: you must set it — a safe pick is 64 / 32 (or 48 / 24, 32 / 16)conv_dim / alpha: same story as Flux (character: 4/1, style: 8/2)Other knobs: repeats, epochs, LR scheduler, optimizer, captions → same playbook as FluxResults: clean, stable, training cost almost identicalSo if you already know how to make a Flux LoRA, you’ll feel right at home. Just remember: Krea is LoRA-only, so don’t forget to set network_dim / alpha.Part 5. Qwen Image LoRA — A slightly different spice ✅Next I tested the Qwen Image base model by building a Belleza LoRA. Honestly, the workflow still felt very familiar — but with a couple of quirks worth noting.Network options: LoRA and DoRA (I stuck with LoRA; DoRA is still new territory for me).network_dim / alpha: defaults to 32 / 32, but I had better luck with 32 / 16.conv_dim / alpha: default is 4 / 4; I dialed it down to 4 / 1 for character LoRA training.Training setup: same recipe as before. For Belleza I used repeat 20 / epoch 5, which trained smoothly. (# image data set = 32) For Momo and Yor Forger LoRAs, epoch 2 was good enough! So I stopped the training. How affordable the training for Qwen is! 🤗Results: just like Flux/Krea — stable, sharp, no extra credit cost.So the takeaway: Qwen doesn’t need you to learn anything brand-new. Think of it as Flux/Krea with slightly different defaults. Adjust network_dim/alpha and conv_dim/alpha to your taste, and you’re good to go.Part 6. Tiny cheat sheet (copy/paste)Character (Flux)conv_dim/alpha: 4/1Repeat/Epoch: ~15 / ~3 (scale to taste)Trigger: unique token at caption start (Keep N=1)Notes: watch visuals > loss; publish best epoch(s)Style (Flux)conv_dim/alpha: 8/2Images: 30–100 (consistent style, 1024×1024)Repeat/Epoch: low (e.g., 10 / 3)Add style hints in captions (“flat colors”, etc.)Multi‑LoRAStart of caption: MainTrigger, SubTriggerShuffle: True, Keep N tokens: 2Flux KreaNetwork: LoRA only → set network_dim/alpha = 64/32 (good default)conv_dim/alpha: same as Flux (character 4/1, style 8/2)Qwen ImageNetwork: LoRA → set network_dim/alpha = 32/16conv_dim/alpha: same as others (character 4/1, style 8/2)Epoch can be low for saturation. Check the progress closely and push the stop button!Part 7. ClosingLoRA making isn’t a strict science—it’s a creative recipe. Pick good ingredients (images), season with captions, adjust heat (params), and taste often (epoch previews). When it looks right, it is right.Have fun, share your models, and let the community riff on them. That feedback loop is where the real magic happens. ✨Good luck & happy training! 🤗

Training AI models might seem complicated, but it’s easier than you think with TensorArt. Here’s a simplified guide to help you get started, even if you’re new.Step 1: Uploading and Managing Your DatasetsUpload Your Images:Accepted formats: PNG, JPG, JPEG.You can upload up to 1000 images for training.Focus on Quality:Use high-resolution images without noise, blur, or watermarks.Enhanced images (like cropped or mirrored ones) can improve results.Delete Images Easily:Click the trash icon on an image to remove it.Step 2: Organizing Your DatasetRegularized Datasets (Optional):Regularization reduces overfitting and helps your model generalize better.If you’re new, skip this step for now.Batch Clipping:Crop your images using tools like Focus Crop (for the main subject) or Center Crop.Recommended sizes:For SD1.5: 512x512 or 768x512For SDXL: 1024x1024 or 768x1024Step 3: Tagging ImagesAutomatic Tagging:Tags are auto-generated when you upload an image.Review and edit these tags for better accuracy.Manual Tagging:Add or adjust tags to match specific traits or features.Batch Tagging:Add tags to multiple images at once. This is handy for trigger words like “seamless pattern” or “holiday design.”Step 4: Setting Training ParametersRepetitions:Decide how many times each image is repeated during training.Choose a Base Model:For 2D characters: AnythingV5 (SD1.5) or Animagine XL (SDXL).For realistic images: EpiCRealism (SD1.5) or Juggernaut XL (SDXL).Advanced Options:Adjust settings like learning rate, epochs, or total steps for more control.Use default settings if you’re unsure.Step 5: Training and TestingStart Training:Training runs one task at a time, so there might be a queue.Schedule during off-peak hours for faster processing.Test Your Model:After training, test your model directly on the workbench.Review preview images to decide whether to retrain or publish.Step 6: Improving Your ResultsRetrain:Not satisfied? Adjust your parameters and try again.Experiment:Try different datasets, models, or settings for better outcomes.ConclusionWith TensorArt, even beginners can train AI models effectively. Start with a simple dataset, use default settings, and gradually explore advanced options as you gain confidence. Take it step by step, and soon, you’ll be creating AI models like a pro!

正确数据集在训练中防止过拟合的重要性在机器学习领域，实现表现良好的模型在很大程度上依赖于训练数据集的质量和适当性。模型训练过程中面临的一个关键挑战是过拟合，即模型过度学习训练数据，包括其噪声和异常值，导致对新数据的泛化能力差。为了减轻过拟合的影响，选择和策划正确的数据集至关重要。以下是防止过拟合的正确数据集的重要性以及如何实现这一目标。理解过拟合过拟合发生在模型变得过于复杂，不仅捕捉到了训练数据中的基础模式，还包括其噪声。这导致在训练数据集上准确率很高，但在验证或测试数据集上表现不佳。实质上，过拟合的模型记住了训练数据，而不是学会了从中泛化。这一问题在数据集过小、噪声大或不具代表性时尤为突出。正确数据集的作用多样性和代表性: 一个好的数据集应该是多样化的，并能代表模型在现实应用中遇到的各种场景。这意味着包括广泛的示例，确保模型学会从不同模式和条件中泛化，而不是记住特定实例。足够的大小: 数据集的大小是一个关键因素。小数据集往往导致过拟合，因为模型没有足够的示例来充分学习基础模式。较大的数据集为模型提供了更多的机会看到不同的示例，减少了过拟合的可能性。平衡和无偏差的数据: 数据集不平衡，即某些类别或条件过多，会导致模型对这些类别产生偏见。这种不平衡会导致过拟合在这些类别上。确保数据集平衡有助于模型更有效地学会在所有类别之间泛化。干净和预处理的数据: 含有错误或无关信息的噪声数据会在训练过程中误导模型。适当的预处理，如去除异常值、归一化值和处理缺失数据，对于提供反映问题域的干净数据至关重要。数据增强技术: 数据增强涉及通过旋转、平移和缩放等变换创建训练数据的变体。这种技术可以人为地增加数据集的大小和多样性，帮助通过向模型展示更多的变异示例来防止过拟合。确保正确数据集的策略交叉验证: 使用交叉验证技术，将数据集分成多个训练和验证集，可以更好地估计模型的表现并帮助识别过拟合。该方法确保模型在不同的数据子集上进行测试，促进更好的泛化。正则化: 应用L1或L2正则化等正则化技术有助于惩罚过于复杂的模型，鼓励生成更简单的、泛化能力更强的模型。这种方法与精心策划的数据集相结合，可以有效防止过拟合。数据划分: 正确地将数据划分为训练集、验证集和测试集是至关重要的。训练集用于训练模型，验证集用于调整超参数，测试集用于评估最终的模型表现。确保这些集合能够代表整个数据集有助于实现平衡的训练过程。监控学习曲线: 通过监控训练和验证损失的学习曲线，实践者可以及早发现过拟合的迹象。如果训练损失持续下降而验证损失开始上升，这就是过拟合的明确迹象。

1.数据集（尽量多图片）和，正则化数据集 （少量图片），（图片尽量只保留要训练的人物，（我训练的图片是游戏立绘和截图，然后抠图只保留要训练的人物））（补充：正则化数据集，没必要可以不上传，仅建议上传头像或全身照，不要有特殊动作！不要有特殊动作！！不要有特殊动作！！！）2.使用底模：我用的是v10 - WAI-NSFW-illustrious-SDXL，训练建模推荐是自己常用的，这方便生成图片，如果生成图片时与训练模型的底模，差异过大，可能无法生成，或（懂了就行）3.触发词（触发词只是方便作图的填写提示词的）：建议只描述发色、瞳孔颜色、面部装饰（耳环、眼镜类）、身体特征（纹身或其他），如果有服装描述，在生成图片的就要删除，（我训练的建模，如陆久的黑羽、fnk的腿环，在不需要时删掉）4.图片处理参数，默认单张重复次数（我保留默认设置20，也可以增减），训练轮数（我训练轮数8-12之间，也可以调整），每 N 轮保存一个5.其他默认设置不变6.样图设置（和平常生成图片一样，就是做个参考）：提示词，对图片的画面要求（下方），加上训练模型人物的外貌和服装特征（触发词加上服装特征，我是这样子的）masterpiece, extremely detailed, highres, detailed beautiful face and eyes, best quality, best aesthetics, perfect anatomy, perfect proportions, high resolution, good colors, bright skin, good shading, counter-shading, well detailed background,, male focus, BREAK,反向提示词，也可以去别人生成图片学习，（lowres, bad quality, worst quality, bad anatomy, sketch, jpeg artifacts, ugly, poorly drawn, censor,blurry, watermark,simple background,）7.训练模型前可以先用提示词，作图模拟

”It's done! I was able to do model training too!”Model training may be asked as an assignment at the event.This is a great reward for the effort, so I definitely want to try it.https://tensor.art/models/806517974249805611/Christmas-Yggdrasil-2024-12-12-07:37:37This is a model I made for an event.All I had to do was have the AI ​​draw three pictures and "munch through" the online training.I will explain my example in 5 steps.First step. Think about the theme of the model. Prepare 3 pictures. Pictures drawn by AI is fine.(Be careful about copyright)It doesn't matter what you want, but if possible, it's best to choose something you like. I like big things, so I chose 3 pieces including this one.Second step: Upload images from the online training page.Then click on the uploaded image and "delete" the inspirational words from the radio buttons on the right. The key is to "erase" rather than "keep".Do this for all images you upload.Third step. Enter parameters. Let me introduce what I think is the easiest and most reasonable method.First, regarding the model, select "SD 3.5 Large" from the "Standard" tab.The reason is because it's cheap now (lol).In fact, model training consumes a lot of credits, so this element is not trivial.（However, the theme for December 18th is to reveal the "illstrious" model, so please be careful about that.）There are several places where you have to type in text.The most important thing is the trigger word. Enter here a short phrase that inspires you.If your only purpose is to create a model, I don't think you need to mess around with the other numbers that much.In my case, "Text Encoder learning rate" was set to 0.00005, "Clip Skip" was set to 2, and "Network Dim" and "Network Alpha" were set to 128.”repeat”was set to 10, and "epoch" was set to 20.Also, LR Scheduler was set to “cosine_with_restarts”.All that remains is to write a sample creation prompt and a negative prompt.This is a prompt for writing a sample, so you can do it just like you would normally do when you ask the AI ​​to draw a picture.And then all you have to do is press the button to start training, and the AI ​​will “study drawing” for you.This amount of learning will take approximately 45 minutes.The Fourth step is to create a project.Create→Add LoRA→Host my model→Create a projectThis will take you to the project launch screen. Enter the necessary information and publish your project. This project will become the ``container'' for the model you created.At this stage, there is no model of the project contents yet.The Final step is to "install" the completed model into the project and publish it.When training is complete, you should see a screen like this: Click "Publish" for the set with the image you like.A “Select Project to Publish” pop-up will appear, so select the project you just created.You will be redirected to the “Host a model” screen, so just enter the model data etc. and publish it.You can perform model training with the above steps. Once you try it, you'll realize it's not that difficult.Learning can be done with AI, but inspiration can only be obtained by humans.Bring your inspiration to life in your models.I hope this article helps you in your creation. Thank you.

Welcome to the world of SDXL LoRA training! If you’re here, you’re probably a 🎨 creative soul itching to dive into the 🎨 magic of custom 🤖 AI models. So, grab your ☕, 🍵, or 🚨 drink of choice and let’s make this journey as 😂 fun and engaging as it is 🔝 informative!The Dataset Adventure: 💎 High-Resolution Gold 💎🖼 Images:Think of your dataset as the 🔲 cornerstone of your 🎨 masterpiece. For SDXL LoRA training, I stick to 30 images at a minimum resolution of 1024 x 1024. But hey, if you can go ⬆️ higher, go for it! More pixels, more 🚀 magic.For modern subjects, 🔄 high-res images are plentiful, but if you’re tackling a niche from the 🕰 90s or 🌐 00s (hello Y2K aesthetics!), brace yourself for some 🛠 pixelated nostalgia. The key is ⭐ quality over 🔹 quantity—crystal-clear images yield crystal-clear ✨ results.🎨 Captioning:Oh, captioning… the unsung 🤟 hero of dataset preparation. I ❤️ swear by WD14 captioning with Kohya-ss—it’s like having a personal 🤖 AI librarian. To supercharge my captions, I use 3️⃣ models:wd-v1-4-convnextv2-tagger-v2wd-v1-4-vit-tagger-v2wd-convnext-tagger-v3Why three, you ask? Because variety is the 🍛 spice of life! These models ensure my 🔗 tags are as descriptive as possible. Plus, the “➕ Append TAGs” feature in Kohya-ss keeps duplicates at 🔬 bay.Got a few tags that feel like party crashers? Enter Booru Dataset Tag Manager (🤝‍♂️ a.k.a. your new best friend). It’s perfect for cleaning up captions. Or, if you’re feeling ✉ old-school, a simple Notepad edit works too. Pro tip: delete responsibly. 📝Training Parameters: The 🌈 Sweet SpotAlright, time to ✌ roll up our sleeves. Training parameters can feel like ⛈ alchemy, but I’ve found a few secrets that work wonders. 🤩2️⃣ Repeats and 5️⃣ Epochs:Repeats: 20 is my magic ✨ number. My dataset folder usually looks something like 20_my_dataset_name.Epochs: Five is the 5️⃣ sweet spot. Anything beyond that tends to 🔄 overtrain. Why wait for 10 epochs when 5️⃣ gets the job done?🌿 Tips for the Road:🎉 Experiment Fearlessly: Training AI models is an 🎨 art and a 🔬 science. Tweak, test, and don’t be afraid to fail forward. 😅💕 Share the Love: Got 🌟 results you’re proud of? Share them! Collaboration is the ❤️ heart of innovation.🔔 Ask for Help: Whether you’re stuck on a ⚠ parameter or looking for feedback, the 🌍 community is your biggest resource.Final Words 🎨Training your own SDXL LoRA is like crafting a custom 🍲 recipe—it’s equal parts 🥤 skill, ⏳ patience, and 🔧 creativity. So go forth, experiment, and most importantly, have fun! 🎩Got questions, 🤔 insights, or just want to share your 🎨 creations? Drop a 🔗 comment or DM me. Let’s make the 🤖 AI art world a little brighter, one LoRA at a time. ✨🌈 Happy training! 🚀BlackPanther

はじめにAI画像生成の急速に進化する分野では、LoRA（低ランク適応）やLoKrなどの手法が、Stable Diffusionのような大規模モデルを微調整する強力な方法として登場しています。これらの手法の違いや、長所と短所を理解し、効果的に適用することは、高品質な画像を効率的に生成しようとする実務者にとって非常に重要です。本記事では、LoRAとLoKrの違いに焦点を当て、それぞれの手法の利点と欠点を探り、LoKr（LyCORISとしても知られる）の詳細な解説を提供します。LoKrがAI画像生成でどのような大きな利点をもたらすかに重点を置きます。LoRAの理解LoRAとは何かLoRA（Low-Rank Adaptation、低ランク適応）は、大規模な事前学習モデルを効率的に微調整するための手法です。モデルのすべてのパラメータを更新するのではなく、学習可能な低ランク行列をモデルのアーキテクチャに注入します。微調整の過程で、LoRAは追加の低ランク重み行列を導入し、タスク固有の情報をキャプチャします。この方法により、更新が必要なパラメータ数が大幅に減少し、計算コストとメモリ要件が低減されます。LoRAの利点効率性：LoRAは少数のパラメータのみを更新するため、微調整に必要な計算資源を削減します。メモリフットプリント：追加の低ランク行列は、完全な微調整と比較してメモリ消費が少ないです。高速性：最適化するパラメータが少ないため、トレーニング時間が短縮されます。LoRAの欠点表現力の限界：低ランク行列では、複雑なパターンを効果的にキャプチャできない場合があります。性能のトレードオフ：場合によっては、LoRAはすべてのパラメータを微調整する方法と比較して、性能がわずかに低下することがあります。LoKr（LyCORIS）の理解LoKrとは何かLoKr（Low-Rank Kronecker product adaptation、低ランククロネッカー積適応）は、適応プロセスにクロネッカー積を組み込むことで、LoRAの原理を拡張した高度な微調整手法です。LoKrは、LyCORIS（Rank-One更新と共有部分空間による低ランク圧縮）フレームワークの一部であり、AI画像生成タスクにおけるモデル適応の効率と効果を向上させることを目的としています。LoKrは、クロネッカー積を活用してより表現力のある適応層を導入し、パラメータ数を大幅に増加させることなく、データ内のより複雑な相互作用やパターンをモデルがキャプチャできるようにします。LoKrの利点表現力の強化：クロネッカー積を使用することで、LoKrはデータ内のより複雑な関係をモデル化できます。パラメータ効率：完全な微調整と比較して、パラメータを大幅に増やすことなく高い性能を実現します。画像品質の向上：特に、AI生成画像の詳細なテクスチャやスタイルをキャプチャするのに効果的です。LoKrの欠点複雑性：クロネッカー積の実装は、適応プロセスに複雑さを加えます。計算コスト：より複雑な操作のため、LoRAよりも計算要求がやや高くなります。LoRAとLoKrの違い適応手法LoRA：モデルの重みに追加される低ランク行列を使用して、タスク固有の情報をキャプチャします。LoKr：クロネッカー積を導入することで、高次の相互作用をモデル化できます。表現力LoRA：低ランク表現の制限により、複雑なパターンのキャプチャが困難な場合があります。LoKr：表現力が強化され、モデルがより複雑なパターンを学習できます。パラメータ効率LoRA：非常にパラメータ効率が高いが、性能を多少犠牲にする可能性があります。LoKr：パラメータ効率と性能のバランスを取り、パラメータを大幅に増やすことなく優れた結果を提供します。計算要件LoRA：必要な計算量が少なく、トレーニングが高速です。LoKr：計算要求はやや高いですが、複雑なタスクでより良い性能を発揮します。LoKr（LyCORIS）がAI画像生成においてもたらす利点1. 優れたディテールのキャプチャLoKrは、画像の細かいディテールをキャプチャするのに優れています。クロネッカー積を活用することで、高品質な画像にしばしば存在する複雑な空間パターンやテクスチャをモデル化できます。これにより、よりリアルで詳細な画像生成が可能になります。2. スタイル転送の向上異なる芸術的スタイルへの適応やスタイル転送を伴うタスクでは、LoKrの強化された表現力により、異なるスタイルのニュアンスをよりよくキャプチャできます。これにより、希望する美的感覚を忠実に再現した画像が生成されます。3. 効率的な適応LoKrは、パラメータ効率と性能のバランスを取ります。すべてのパラメータを更新する必要なく、モデルを新しいタスクに微調整でき、計算資源を節約しながら高品質な結果を提供します。4. 柔軟性この手法は、モデル内のさまざまな層に適用でき、適応がどのように、どこで行われるかの柔軟性を提供します。これにより、実務者はタスクの具体的なニーズに合わせて微調整プロセスをカスタマイズできます。LoKr（LyCORIS）の詳細な解説インストールや実際の操作手順は扱いませんが、LoKrの動作を深く理解することで、実務者はその使用について適切な判断を下すことができます。LoKrにおけるクロネッカー積クロネッカー積は、2つの小さな行列からブロック行列を生成する数学的な操作です。LoKrの文脈では、パラメータ数を大幅に増やすことなく、高次の相互作用をモデル化できる適応行列を作成することが可能です。クロネッカー積を利用することで、LoKrはモデルの層により表現力のある変換を注入できます。これにより、モデルはデータ内の複雑な関係を学習でき、特に複雑なパターンやテクスチャのキャプチャが必要な画像生成タスクに有益です。パラメータ効率と性能LoKrは、パラメータ数とモデルの性能とのバランスを維持します。クロネッカー積を用いて適応行列を慎重に設計することで、大量の追加パラメータを必要とせずに、表現力を向上させています。この効率性は、計算資源が限られているが高い性能が求められる状況で特に重要です。Stable Diffusionへの適用性LoKrは、Stable Diffusionモデルの微調整に特に適しています。新しいスタイルや主題への効果的な適応により、高品質な画像を生成するモデルの能力を高めます。LoKrの柔軟性により、モデルのさまざまな部分に統合でき、AI画像生成分野の実務者にとって強力なツールとなります。結論AI画像生成の分野では、LoRAとLoKrの両方が、大規模モデルを効率的に微調整するための有用な手法を提供します。LoRAはシンプルで資源効率の高いアプローチを提供しますが、LoKr（LyCORIS）はクロネッカー積を導入することで、より複雑なパターンや相互作用をキャプチャする能力を拡張しています。LoKrは、画像品質の向上、詳細なテクスチャのキャプチャ、新しいスタイルへの高忠実度な適応において際立っています。その利点により、AI生成画像の可能性を広げようとする実務者にとって、魅力的な選択肢となっています。LoRAとLoKrの違いを理解し、LoKrがAI画像生成において持つ強みを認識することで、実務者は自分のニーズに最も適した手法を選択することができます。

LoRA (Low-Rank Adaptation) is an add-on commonly used by AI Arts users to refine artworks generated from Stable Diffusion Model Checkpoints according to their preferences.This small additional model can apply impressive changes to standard Model CheckPoints with relatively good quality, depending on its processing.LoRA models typically have a capacity between 10 – 200 MB, which is significantly smaller than checkpoint files that exceed 1 GB.However, it's important to note that various factors contribute to the varying sizes of LoRA, such as its settings/parameters. Nonetheless, LoRA generally has a smaller capacity than Model Checkpoints.Given its relatively small size, LoRA makes it easier for users to upload, download, and combine it with the checkpoint models they use, enriching the AI Arts results with backgrounds, characters, and styles according to the user's desires.It's important to note that LoRA cannot be used independently. It requires a checkpoint model to be functional. So, how do you create a LoRA? What do you need to prepare to make a LoRA?To create a LoRA, you need to prepare the following:Datasets Before creating a LoRA, ensure you have datasets that will be trained/used to create the LoRA. If you don't have one, you must prepare it first.There is no set rule for the quantity. Too many datasets don't necessarily yield a good LoRA, and too few may not either. Ensure to prepare high-quality datasets with various angles, poses, expressions, positions, and others. Align it with your goal in creating the LoRA, whether it's for style, creating fictional/realistic characters, or other purposes.Often, users process their datasets first before training them to become LoRA. This process includes cropping to ensure all datasets are the same size and adjusting the image resolution for better quality and clarity. If the datasets are blurry, the LoRA might produce poor and blurry quality.Understanding LoRA Parameters and Settings Understanding LoRA parameters and settings is not easy, but you can learn it gradually over time if you are determined to study it. This is necessary knowledge that you must have.There might not be many articles in Indonesian discussing LoRA, so to learn more, you may need to dive deeper into various English or other language sites that cover the topic, join the community, and learn what they are learning too.A commonly used formula for LoRA is "Datasets x num_repeats x epoch / train_batch_size = steps", which calculates the number of steps for LoRA and sets the steps for each epoch.Example: Datasets (40 images) x 10 (num_repeats) x 10 epochs / 4 (train batch size) results in 1000 steps for the 10th epoch of LoRA. Steps for the first epoch are 100 steps, the second epoch: 200 steps, and so on until the 10th epoch, which has 1000 steps.This formula is not the only thing you need to know. You also need to learn about network_dim, network_alpha, learning_rate, and many others.PC/Colab The most crucial and primary thing you need to have is a PC with Stable Diffusion or Kohya Trainer installed locally. However, this requires a high-spec computer, including a powerful processor, RAM, high VRAM GPU, and others. If you don't have one, an alternative is to train LoRA using Google Colab.Luckily, now you can easily train Lora in tensor.art, by logging in to your account, clicking on profile then selecting train Lora, or you can click here to go directly to the train lora pageLocal/Cloud Storage Media Make sure you have storage media to save your datasets and LoRA. You can upload these datasets and LoRA to popular cloud storage services like Google Drive, Mega, MediaFire, etc., so if your files are lost from your computer, you still have a backup.You can also upload them to platforms like HuggingFace, Civitai, and others. This will make it easier for you to use them through Google Colab.

How to Train LORAs on Tensor.art for Realistic and Cartoon Images: A Complete Guide to Prompts and ParametersIntroduction: The Power of Tensor.art and LORAsTensor.art has established itself as an accessible and powerful platform for generative AI enthusiasts and professionals, especially for creating realistic and stylized images. Among its most versatile tools are LORAs (Low-Rank Adaptation Models), which allow users to adapt pre-trained models (like Stable Diffusion) to generate customized content, from hyper-realistic portraits to unique cartoon characters.In this article, we’ll explore how to train LORAs on Tensor.art and how to maximize parameters, codes, and prompt engineering techniques to achieve precise and creative results.Part 1: Understanding LORAs and Their Role in Image GenerationWhat Are LORAs?LORAs are lightweight adaptations of existing AI models, designed to add specific layers of learning without requiring a full retraining of the base model. This means you can "teach" the model new concepts (e.g., an art style or character) with minimal computational resources.On Tensor.art, LORAs are ideal for:Creating consistent faces or characters across multiple images.Developing unique styles (e.g., "clean line" cartoons or photographic realism).Tailoring models to specific needs (e.g., clothing, settings).Part 2: Training Your Own LORA on Tensor.artStep 1: Preparing the DatasetTraining quality depends directly on your dataset. Follow these guidelines:For Realism: Use high-resolution photos with varied lighting and angles. Include close-ups and full-body shots.For Cartoon: Collect images with consistent linework (e.g., bold outlines, flat colors) and avoid abrupt style variations.Ideal Quantity: 50–200 images, depending on the theme’s complexity.Step 2: Configuring Training on Tensor.artTensor.art simplifies LORA training:Navigate to "Train Model" and upload your dataset.Set parameters:Epochs: 50–150 (avoid overfitting).Batch Size: 2–4 for basic GPUs.Learning Rate: 0.0001–0.0002 to balance speed and precision.Add descriptive tags to images (e.g., "blue_eyes", "anime_style") to link concepts to the model.Step 3: Fine-Tuning for Realism vs. CartoonRealism: Enable "High-Res Fix" and use embeddings like RealisticVision for skin textures and details.Cartoon: Add style triggers to prompts (e.g., "makoto shinkai style") and lower cfg_scale (5–7) for artistic flexibility.Part 3: Mastering Parameters and Prompts for PrecisionKey Technical Parameters on Tensor.artCFG Scale (7–12): Controls adherence to the prompt. Higher values (12+) suit realism; lower values (5–7) favor stylization.Sampler: Use DPM++ 2M Karras for realism and Euler a for cartoons.Steps (30–50): More steps enhance details but increase generation time.Crafting an Effective PromptA well-structured prompt blends technical and descriptive elements. Example for realism:RAW photo, (a detailed portrait of a woman:1.3), (piercing green eyes:1.2), soft natural lighting, skin pores, (cinematic depth of field:0.9), Nikon D850, 85mm lens Negative prompt: cartoon, blurry, deformed, low-res Key Elements:Weighting with Parentheses: (element:1.3) increases emphasis; (element:0.8) reduces it.Specific Details: Mention cameras, lenses, and lighting to reinforce realism.Negative Prompts: Block unwanted styles (e.g., "3D render", "anime").Example for Cartoon:Studio Ghibli style, (a cheerful boy with spiky hair:1.4), vibrant colors, magical forest background, cel-shading, (soft gradients:0.7), by Hayao Miyazaki Negative prompt: realism, photorealistic, noise Part 4: Combining Multiple LORAs and ModelsOn Tensor.art, you can merge LORAs for complex results. For example:Combine a facial realism LORA with a dramatic lighting LORA.Use a base model like DreamShaper for cartoons and add a water texture LORA.Best Practices:Test combinations with adjusted weights (e.g., <lora:lighting:0.7>).Avoid conflicting styles (e.g., realism + cartoon).Part 5: Common Mistakes and How to Fix ThemOverfitting: Repetitive or low-variation images.Solution: Reduce epochs and diversify the dataset.Lack of Detail:Solution: Increase cfg_scale and add descriptors like "4K", "ultra-detailed".Style Inconsistency:Solution: Use unique triggers (e.g., "my_style_v1") and reference them in prompts.Conclusion: Elevating Your Art to the Next LevelTraining LORAs on Tensor.art is a journey of technical and creative experimentation. By mastering parameters, prompts, and dataset curation, you can create everything from photorealistic portraits to immersive cartoon worlds. Remember: The key lies in constant iteration and meticulous analysis of results.Additional Resources:Experiment with hybrid prompts (e.g., "realism with surreal touches").Join the Tensor.art community to share LORAs and tips.Document your tests in a notebook for future refinement.With practice and attention to detail, your Tensor.art creations will reach professional standards, whether for personal or commercial projects.

In this guide, we’ll walk through the full process of online training on TensorArt using Wan2.2. For this demo, we’ll be using image2video training so you can see direct results.Step 1 – Open Online TrainingGo to the Online Training page.Here, you can choose between Text2Video or Image2Video.👉 For this tutorial, we’ll select Image2Video.Step 2 – Upload Training DataUpload the materials you want to train on.You can upload them one by one.Or, if you’ve prepared everything locally, just zip the files and upload the package.Step 3 – Adjust ParametersOnce the data is uploaded, you’ll see the parameter panel on the right.💡 Tip: If you’re training with video clips, keep them around 5 seconds for the best results.Step 4 – Set Prompts & Preview FramesThe prompt field defines what kind of results you’ll see during and after training.As training progresses, you’ll see epoch previews. This helps you decide which version of the model looks best.For image-to-video LoRA training, you can also set the first frame of the preview video.Step 5 – Start TrainingClick Start Training once your setup is ready.When training completes, each epoch will generate a preview video.You can then review these previews and publish the epoch that delivers the best result.Step 6 – Publish Your ModelAfter publishing, wait a few minutes and your Wan2.2 LoRA model will be ready to use.Recommended Training Parameters (Balanced Quality)Network Module: LoRABase Model: Wan2.2 – i2v-high-noise-a14bTrigger words: (use a unique short tag, e.g. your_project_tag*)*Image Processing ParametersRepeat: 1Epoch: 12Save Every N Epochs: 1–2Video Processing ParametersFrame Samples: 16Target Frames: 20Training ParametersSeed: –Clip Skip: –Text Encoder LR: 1e-5UNet LR: 8e-5 (lower than 1e-4 for more stability)LR Scheduler: cosine (warmup 100 steps if available)Optimizer: AdamW8bitNetwork Dim: 64Network Alpha: 32Gradient Accumulation Steps: 2 (use 1 if VRAM is limited)Label ParametersShuffle caption: –Keep n tokens: –Advanced ParametersNoise offset: 0.025–0.03 (recommended 0.03)Multires noise discount: 0.1Multires noise iterations: 10conv_dim: –conv_alpha: –Batch Size: 1–2 (depending on VRAM)Video Length: 2Sample Image SettingsSampler: eulerPrompt (example):TipsKeep training videos around ~5 seconds for best results.Use a consistent dataset (lighting, framing, style) to avoid drift.If previews show overfitting (blurry details, jitter), lower UNet LR to 6e-5 or reduce Epochs to 10.For stronger style binding: increase Network Dim → 96 and Alpha → 64, while lowering UNet LR → 6e-5.

Hello and welcome to the fourth edition of REYApping, a space where I write a bunch of nonsense. Without further ado, let's begin.Fourth edition? Dang, I didn't expect I'd make a fourth edition of me yapping, but here I am, trying to yap about my experience with model training, with the model this time being "Illustrious". Now I warn you that I never use this in my entire life before. Even if it's based on XL, when trying it, it just felt very different from Animagine or other anime XL in general. Also quick disclaimer: any settings that I used here is just an uneducated guess, so I encourage you to NOT STRAIGHT UP USE THESE SETTINGS. Unless it's coincidentally good. Without further ado, let's begin.Step 1, Getting IdeasI'm not really a creative person, so I admit that this part is one of, if not the hardest part. Currently I got a task telling me to train with Illustrious as base, put it in either visual, game, or space design, and get it into Tenstar Fund. The required channel reminds me of the past Hunyuan event with basically the same task except the Tenstar Fund thing, and I made: Fumo Dolls.I decided to revisit the image that I gathered and see some really good opportunity. Most of my dataset contains anime and/or manga characters with 90% of it being Tohou characters, and since we're dealing with Illustrious, everything just aligned well, so I can continue to the next step.Step 2, Dataset Gathering and CaptioningSince I choose to use my old images, I can skip the gathering process, but if you started from scratch, then I recommend you to either generate at least 1MP image, or if you look for images in Google, you need to use the advanced search and pick the "larger than 2MP" in size. If you have an interesting concept but don't know where to look and Google is just a no go, then I recommend to try generating with Dall E that you can access through Microsoft Edge's Bing. That will get you 1024x1024 images, which is good enough for training. For this Fumo model, I use 15 images of real Fumo dolls, with different backgrounds, angles, and subjects. After getting the images you want, it's time to do the most annoying part: captioning.Captioning images is basically telling the base model what the image is all about. If you train a style, then you need to caption pretty much everything in the image. For a character, you assign its key characteristics into a unique word (called trigger word) and just caption its clothes and what it is doing. There're other methods, but I'll be captioning everything in this one. You need to create a txt file that corresponds to the image name. To make it easier, I renamed my image to "fumo" and let Windows assign a number. I then create txt files and renamed it the same. The result pretty much look like this:Since the dataset mostly contains characters that is recognized by Illustrious, I'll be captioning it by using the trigger word first, character name, and then clothing (if different from original), pose, and then background. After it's done, I zipped all and it's ready to be uploaded.Step 3, Training parametersI'll just post a picture off my parameters here.Also, since the training image is real photo, I add "realistic" tag at the end of all images by doing this:Now it's time to train.Part 4, Test and PublishAfter the training is done, I published 2 models (Epoch 3 and 6) for tests since Tensor can't use an unpublished model for testing. The result is quite okay.It apparently also creates cute randomness like this:Final ThoughtsThis is quite interesting. Illustrious as a model seems very good, it can make nice images and understands a lot of characters so it makes it easier for people to create character images without LoRas. This one will be the first of many tests that I will do with this base model. Also, I'll make an update later on with some more base models (like SD3.5 perhaps?).Thank you for reading this part of REYApping. See you in the next one.

The number of steps and images required to generate a checkpoint in Tensor Art depends on several factors, including your model architecture, the complexity of the task, and the quality of the data. Here's a breakdown to help you estimate:1. Number of StepsThe required number of steps depends on:Dataset Size: Larger datasets need more steps for sufficient training.Learning Rate and Convergence: Smaller learning rates typically require more steps for the model to converge.Task Complexity: Complex tasks (e.g., image generation, multi-class classification) need more training steps than simpler tasks.General Guidelines:Small Dataset (e.g., 1,000 images): 1,000–5,000 steps.Medium Dataset (e.g., 10,000–50,000 images): 10,000–50,000 steps.Large Dataset (e.g., >100,000 images): 50,000+ steps, often with early stopping to prevent overfitting.2. Number of ImagesFor generating a meaningful checkpoint:The model typically needs at least 1,000–10,000 diverse images for tasks like image generation or classification.For high-quality results, datasets like COCO (Common Objects in Context) or ImageNet often include 50,000+ images.If you're working with custom data:Aim for a minimum of 1,000 images for fine-tuning pre-trained models.If training from scratch, 10,000–50,000 images is a good starting point for robust model performance.3. When to Create CheckpointsCheckpoints are typically saved during training:After each epoch (one pass through the dataset).At regular intervals (e.g., every 1,000 steps).Based on validation performance, to save the best-performing model.Example WorkflowIf you have 10,000 images:Set up training for 20,000 steps (2 epochs if batch size = 32).Save checkpoints every 1,000 steps or at the end of each epoch.Evaluate the model after each checkpoint to decide if further training is necessary.Key TakeawaySteps: 1,000–50,000+ depending on task and dataset size.Images: 1,000+ (fine-tuning) or 10,000+ (training from scratch).Checkpoints: Save at regular intervals to monitor progress and ensure you don't lose training data in case of interruptions.

首先点击右上角的头像，在弹出的下拉框中选择我训练的模型，进入训练中心。如果之前有训练过模型，这里会看到许多训练任务。然后选择在线训练按钮进行一次训练。左侧是数据集窗口，默认没有任何数据。您可以上传一些图片作为数据集，或者上传一个数据集压缩包，压缩包可以包含标注文件，格式和kohya-ss一样，每个图片文件对应一个同名的标注文件txt。右边的模型主题中可以选择二次元人物、真实人物、2.5D、标准以及自定义。这里我们选择自定义，在使用底模中选择SD3模型。注意在选择版本中下拉框内选择T5XXL的版本，这样才可以训练T5文本编码器。基础模式下参数选择，推荐单张图片重复次数4，轮数为16。上传一个处理好的数据集后，如果你的数据集标注中有人物名，可以不写触发词。否则你应该给你的模型起一个简单的触发词，例如人物名称或者风格名称。接着从数据集中选择一个标注文件作为预览提示词。如果你想使用专业模式，选择右上角按钮切换到专业模式。专业模式推荐学习率翻倍，然后使用cosine_with_restarts学习率调度器，优化器可以选择AdamW8bit。开启打乱标签(shuffle)，并且保持第1个token(如果你有一个人名触发词在第一个)关闭噪声偏移功能，卷积DIM和Alpha可以选择8和1。在样图设置中追加填写反向提示词，接下来就可以开始训练了。在训练队列中，你可以看到当前loss值变化表以及每轮epoch产生的4张样图。最后可以选择效果最好的epoch下载到本地或者直接在tensorart上发布。

So you want to make a LORA, but aren't sure where to start? Hopefully, after reading these articles, you'll decide this is something you'd like to try. I've tried to include examples and helpful resource links.ADDENDUM: Since writing this there have been (or will be by the time you read this) changes here at tensor.art. As I write this what those changes may entail, how they may affect TOS policy or LORA creation is unclear. However, in a effort to get ahead of things, I wrote a follow-up article about FACSIMILE images, at least how I define them, and what that means as many of my test LORAs used such AI generated images. The article can be found here: https://tensor.art/articles/884619347480304594 If you’re like me you have probably found yourself sitting there after your image generated, staring at the screen, thinking: ‘I don't like how this looks.’ Wishing you had a better LORA to use for your artwork. Or maybe you were muttering something with more colorful metaphors because, well, things and stuff happens.Well, curse at your screens no longer! Making a LORA, at least here on Tensor.art, is surprisingly simple.No. Really. If I can do it so can you.There are many how to guides for LORA creation that will take you step-by-step through the process. This article isn’t one of those. This offers some practical advice and, hopefully, useful information based on my own observations (good, bad, and d’oh!) from creating a couple of passable LORAs.To begin have a basic idea of the what and why of your goal for creating a LORA. Even if it’s just to see how it’s done, you need a good basic thematic end result in mind. Meaning what type of character or art style you'd like to create with your LORA. Anything is possible from Dark Jedi, Barbarian Warrior, Frank Thorne art style, Pies, race cars, Whimsical Kittens, Post-apocalyptic Landscape, and etcetera et al.Main question: What kind of art are we wanting to create?Not sure. Then ask yourself: What character or style do I need (or want to create) that there isn’t already a LORA for. Or that no LORA exists for your preferred generative Model. Or that you just would like a better version of. (Not all LORAs are created equal. You can try three similar character LORAs and often get very different results.)For example there are many Character LORAs for actress Kaley Cuoco. But there were none of her as a fantasy warrior. Of course with the right LORA combos you can create awesome Kaley Cuoco Barbarian Warrior artwork. But what about a blood-splattered Warrior? Sure, we can create that, but using how many LORAs to get the image just right?I took this idea and created BARBARIAN WARRIOR KALEY using 25 images in the training dataset. The images I curated include 7 head and shoulder shots, 3 mid close-up body shots (head to knees), 10 full body shots in dynamic poses, 1 artistic image, and 4 sample images using a test prompt similar to the one used to generate the LORA sample images generated during training then in-painting to touch up the face and eyes.Here’s the published LORA link: https://tensor.art/models/876777897912706295?source_id=njeyo1nrnEW3oPYsaX309xkk so you can view the gallery of generated images. This LORA appears to do dungeon backgrounds and blood-splatter very well. In fact it may seem too bloody, depending on your prompt, and not enough like Kaley even at the default weight of 1. Lowering the weight to .8 lessens the persistence of blood-splatter, but this may affect the likeness since the source images were AI generated facsimiles. But it's good enough we can use it with a Character LORA, if we want a more realistic likeness.Now, having seen some of what is possible, all you need is to…Step 1. Decide what type of LORA you want to create, and what you want it to do for you.(PART 2 HERE: https://tensor.art/articles/876855087064988656)