Welcome To Tensor! So you just started your journey in AI generation, now what?
This Guide is what I wished I've known when I started Tensor. I'll share everything I've learned so far (•̀ᴗ•́ )و
This Guide will not give you a detailed explanation on how AI work, it's meant to help you get accustomed to Tensor and with generating art.
Table Of Content
Common Terminology
Create Tool (Text2Image)
Generation basics
Prompts In Depth
Recommendations
Creations settings
Checkpoints
Lora
Embedding
Artist Tags
Advanced Generation tips
Other Create Tools
Exploring Tensor
Advice
Terminology
Checkpoint: It's the main model that will be used to draw. Consider this your cake. it's the base of your creation.
LoRA (Low-Rank Adaptation): This is your topping, your decoration, that will distinguish your cakes from the rest.
Embedding: They help the generation either focus on certain elements, or avoid deformity. consider it the oven that automatically turns off when your cake is baked, ensuring it doesn't burn. consider it a cheat sheet for the AI.
Settings
Sampler: How the AI denoises the image. In cake terms, the AI starts with all the ingredients, the sampler is the steps it takes to mix and bake the cake you desire.
Schedulers: control how fast and in what order the AI removes noise (the steps it takes to bake)
Sampling steps: The number of times the AI "refines" the image from pure noise to your final result. higher steps = better results (will discuss it more in recommendations)
CFG Scale (Classifier-Free Guidance): How strictly the AI follows your text prompt vs. getting "creative."
Seed: A number controlling randomness. Use the same seed if you want similar results.
Clip Skip: CLIP is the part of Stable Diffusion that converts your text prompt into numerical data the AI understands, 1: means it uses all layers, the higher the value the more layers the AI will skip, causing more abstract creative generations.
Upscale
Increases the resolution of your creation, can sometimes help fix deformities
Not always necessary, lower upscalers like 1x can sometimes make the image look muddy, if you use Lora
Comparsion of upscaler results using 4X_foolhardy_remacri

Adetailer
Clean up the generation, mostly fixing deformity in the face and hands.
Even if there's no deformity, it can be used to improve the quality of the image aesthatically
Using the face_yolov8m.pt. The ruffles of the skirt were slightly tweaked, and the face was changed (if it's an improvement or not depends on your taste)


Layer Diffusion
To make a transparent background
NOW ENOUGH TERMS Lets go to the fun stuff!
Create Tool (Text2image)
Prompt: What you want the AI to generate [Figure 1]
Negative prompt: what you want the AI to avoid generating. (it's not really a necessary to write in it) [Figure 2]
A1111: I'm still not fully sure, something about stable diffusion, and cloud. what you need to know if you use a lot of emphasis "(), ():1.4 etc" you will be asked to either tweak your prompt or tick the A1111 [Figure 3]
Translation: Translate your prompt to English (quite self explanatory) [Figure 4]
Random: Generates you a random prompt [Figure 5]
Abstract (VERY IMPORTANT): Verbalizes a picture you share with it. with Booru tags, or Natural Language [Figure 6]
Prompt Enhance: Enhances your prompt, it removes artist prompts, some of the positive prompts. Tread carefully with using it. [Figure 7]
Reset: Resetting all settings back to default [Figure 8]
Presets: Saving settings, checkpoints, Lora, Prompt, negative, everything. in-case you want to replicate it, [Figure 9]

Generation Basics [Let's begin with the fun! •̀ω•́ ]
Pick a Checkpoint (I'll recommend ones in the the next part =w=)
Loras for style, specific character, details, poses, props etc.
Embedding (optional) to reduce deformity
Negative Prompt (optional doesn't always work)
The negative prompt I use "score_4, score_5, score_6, modern, recent, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, sketch, comic, meme, poorly drawn, simple background, watercolor, lowres, low detail, text, blurry, realistic, 3d, bad anatomy, bad proportions, deformed anatomy, deformed face, deformed eyes, monochome, grayscale, text, twitter_username, artist_twitter, anatomically incorrect hands missing fingers, extra digits, fewer digits, bad eye, shota, censor, (multiple fingers), (blurry eyes), (extra legs), conjoined,ai-generated, crop out, white line, stubble"
Most checkpoints have a recommendation that you can apply, which includes the negative.
Prompt (The fun part =w=)
Wording Anime Vs Realism
If your gonna use realistic or FLUX models: You'll have to describe the images like you'll describe a photograph, using natural language.
'Full body photograph of a man with black hair sitting on bed resting his arms behind his head."
For Anime [my specialty ( • ̀ω•́ )] and semi-realism models: Use Booru tags (using trigger words.)
"1boy, Black hair, sitting, ,arms_above_head, on bed"
Cheat Sheet =w=: Use the abstract tool to help you verbalize a picture, granted you have to provide the picture, let it be a picture you generated and would like to replicate, or an inspiration (You can only use this tool for SFW pictures only)
Example Of How Abstract words my generation
My prompt and resulting picture

Now how abstract worded it:

Prompt In depth
Format
If you are gonna connect two words you can do it as. "bowl_cut" instead of "bowl cut" it sometimes help the AI understand the prompt better.
When generating smut (Don't worry I don't judge~) Write "rating_explicit, NSFW, Uncensored" at the start of the prompt, it'll help combat any censorship the AI will try.
Separate paragraphs by using "BREAK," (not really necessary IMO)
doesn't work with A1111
Emphasis
Every word in the prompt has a value of 1, if you want to emphasize a word you either use "()", or "(:1.4)", and if you want to de-emphasize just do (blue eyes:0.9), any value less than 1.
Every "()" increases the emphasis by 10% to a max of 30% "((()))" so if you want to emphasize blue eyes either use "(((blue eyes)))" or "(blue eyes:1.3)"
if you want to use "()" without emphasizing just write it as "\(blue eyes\)"
"[ ]"
Use it when the AI doesn't generate the thing you want. Let it be making the subject a girl instead of a boy. making their shirt a different color than the one you wrote, however emphasizing too much can cause the AI to neglect other aspects, which may result in deformity.

Artist Prompt
Writing an artist name in your prompt can alter your creation. You can write it as
(artist:rinotuna), or as rinotuna
Some artists the AI doesn't recognize. You'll have to use Lora to for their style.
tip If you are writing a bunch of artists, just like In the figure I showed previously regarding abstract tool, it'll be good to lower their values, so they don't clash.
Here's an example of how adding an artist name in the prompt can change the style.
Positive Prompt
Positive keywords, used to improve the end result.
Like "Masterpiece, Highres, Absurdres, best aesthatic, high resolution, amazing quality, 4k, hd, ,very aesthetic, volumetric lighting, perfect lighting, detailed eyes, perfect anatomy, perfect proportions, high definition, best quality, very awa, newest, extremely detailed, highres,detailed beautiful face and eyes,best aesthetic, scenery, good colors, maximalism, (Detailed background), Dynamic lighting, deep shadows, dynamic shadows, countershading, depth of field, ultra-detailed,ambient occlusion, raytracing, HD eyes"
it's unnecessary from testing, I've noticed that adding them, doesn't change the outcome of the gen.
Recommendations
Creation Tool (foundation, it's boring but necessary)
Preferred Aspect Ratios beside the recommended (square, landscape, portrait)
512, 768 - 521, 1536 - 710, 1536
768, 1536 - 1024, 1536 - 1200, 1536, 1536, 1080
Samplers (a is for ancestral):
Euler a: balanced in speed and quality
DPM ++ 2S a Karras: More stable than DPM ++ 2M not as high in quality
DPM ++ 2M Karras: Slower, High Quality

Sampling Steps:
Anime, Semi-realism: 20-25
Realistic, Flux Realism: 30+ (30-35 is a good median)
CFG, Guidance Scale:
4 - 7
most checkpoints have it in the recommended setting
Clip Skip:
1-2 (Personally I go for 6)
most checkpoints have it in the recommended setting
Upscaler Models:
4X_ultrasharp basic but reliable
4X_foolhardy_remacri (personal favorite)
Esragan 4X+ :The original, can have better results than the improved version
R-Esragan 4X+ : Improved version of Esragan best for Realistic Photos
R-Esragan 4X+ Anime6B: Best For Anime, Cartoon-ish gens
Here is an Upscaler comparison:



Upscaler Settings
Steps 20-30 don't make it more than your sampling steps
Denoising strength 0.4 - 0.5 feel free to test it out and find the ones you like
Denoising means there will be redrawing. It may cause deformity especially with 1.5x+
if you want to higher res your picture without that much denoising, set the strength to 0.1
Here's an example with 1.5X Upscale

0.1 denoising strength improved the quality of the picture but didn't redraw much over it.
0.4 denoising strength improved the quality of the picture more significantly, however it also caused deformity, the eyes quality improved but now the hands seem to have an extra finger.
Adetailer Models:
face_yolov8s.pt
face_yolov8m.pt
Adetailer Settings:
Steps 20-25
Detection model:
0.5 is a perfect median
low value=high range of repair / High value= High Repair Accuracy
It depends on your image, If there's plenty of deformity across the picture then use 0.3-0.5
If the deformity and mostly in a specific place, like the face. Then use 0.5-0.75
Denoising Strength: 0.4-0.5 (I personally put it 0.25 if I want that unfinished look to the image)
Inpain mask blur: it's for the inpaint tool (Which I don't play with, I'll update after playing with it ദ്ദി ˉ͈̀꒳ˉ͈́ ))
WITH THAT we are done with technical Jargon now for the fun part ⸜(。˃ ᵕ ˂ )⸝♡~
In the next part [Sorry but I've reached the max letter count (╥﹏╥)]
Checkpoint, Lora, Embedding, Artist tags would be their own article.