Guide to Using SDXL

I occasionally see posts about difficulties in generating images successfully, so here is an introduction to the basic setup.

1. Introduction

SDXL is a model that can generate images with higher accuracy compared to SD1.5.

It produces high-quality representations of human bodies and structures, with fewer distortions and more realistic fine details, textures, and shadows.

With SD1.5, generation parameters were generally applicable across different models, so there was no need for specific adjustments.

However, while SDXL can still use some SD1.5 techniques without issues, the recommended generation parameters vary significantly depending on the model.

Additionally, LoRA and Embeddings (such as EasyNegative) are completely incompatible, requiring a review of prompt construction.

Notably, embeddings commonly used in SD1.5 negative prompts are recognized merely as strings in the XL model, so you must replace them with corresponding embeddings or add appropriate tags.

This guide explains the recommended parameter settings for using SDXL.

2. Basic Parameters

VAE

Selecting "sdxl-vae-fp16-fix.safetensors" will suffice.

Many models have this built-in, so specification might not be necessary.

Image Size

Using the presets provided by TensorArt for resolution should be sufficient.

Small or excessively large resolutions may not yield appropriate generation results, so please avoid using the sizes that were frequently used with SD1.5 wherever possible.

Even if you want to create vertically or horizontally elongated images, do so within the range that does not significantly alter the total pixel count (adjust by increasing height and decreasing width, for example).

For example, 1152x896, 1216x832, and 1344x768 are often used.

Sampling Method

Choose the sampler recommended for the model first.

Then, select according to your preference.

Typically, selecting Euler a or DPM++ 2M SDE Karras should work well.

Sampling Steps

XL models might generate images effectively with lower steps due to optimizations like LCM or Turbo.

Be sure to check the recommended values for the selected model.

CFG Scale

This varies by model, so check the recommended values.

Typically, the range is around 2 to 8.

Hires.fix

For free users, specifying 1.5x might hit the upper limit, so use custom settings with the following resolutions:

768x1152 -> 1024x1536

1152x768 -> 1536x1024

1024x1024 -> 1248x1248

Choose the upscaler according to your preference.

Set the denoising strength to around 0.3 to 0.4.

3. Prompt

SDXL handles natural language better.

You can input elements separated by commas or simply write a complete sentence in English, and it will generate images as intended.

Using a tool like ChatGPT to create prompts can also be beneficial.

However, depending on how the model was additionally trained, it might be better to use existing tags.

Furthermore, some models have tags specified to enhance quality, so always check the model’s page.

For example:

AnimagineXL3.1: masterpiece, best quality, very aesthetic, absurdres is recommended.

Pony Models: score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up is recommended.

ToxicEchoXL: masterpiece, best quality, aesthetic is recommended.

In this way, especially for XL models, particularly anime or illustration models, appropriate tag usage is crucial.

4. Negative Prompts

Forget the negative prompts used in SD1.5. "EasyNegative" is just a string.

The embeddings usable on TensorArt are negativeXL_D and unaestheticXLv13.

Choose according to your preference.

Some models have recommended prompts listed.

For AnimagineXL

nsfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]

For ToxicEchoXL

nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digits, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name.

For photo models, sometimes it is better not to use negative prompts to create a certain atmosphere, so try various approaches.

5. Recommended SDXL model

ToxicEchoXL

https://tensor.art/models/689378702666043553

1girl, gray hair,black witch clothes, earrings, pointy ears, elf, veil, cleavage , moon, own hands together, night, seraphic Candles, upper body, looking away, masterpiece, best quality, aesthetic

This is an ultra-high performance model specialized for illustrations.
It has a rather unique painting style because of its original learning and adjustment based on watercolor.

ToxicEchoIL

https://tensor.art/models/810855539031577236

$1girl, kayoko $blue archive$, white white background, standing, upper body, black hoodie, looking at viewer, very awa, masterpiece, best quality, highres, absurdres, newest, aesthetic,$

Super high performance model based on Illustriou base with similar learning and adjustments as XL.
Strong in art style, composition, character and NSFW, most recommended model.

Minimalism Illustrious

https://tensor.art/models/816324441148646205

$1girl, kayoko $blue archive$, deformed, chibi, polka dot background$

Based on Illustrious, this model specializes in minimalism.

Realistic Illustrious Photography

https://tensor.art/models/818074232299784473

$1girl, ui $blue archive$, blue archive, cowboy shot, sitting, asian girl, indoors, library, book stack, reading , depth of field, bokeh, detailed, noise, masterpiece, best quality, newest, highres, absurdres, realistic, photorealistic , film grain, cinematic still,$

Realisticized model based on Illustrious.
It excels at film-like colors, bokeh, and Asian people.
Since it is based on Illustrious, it is also strong with characters, making it easy to create cosplay-style photos.

6. おわりに

I hope this will help those who are having trouble generating it.

I also create various LoRAs to change the style, so please visit my user page.

https://tensor.art/u/649265516304702656

I am planning to make a new model when the successor model is released.

SDXLモデルの利用手引

ここではSDXLの基本的な設定を紹介します。

1. はじめに

SDXLはSD1.5と比較してより高精度な生成が行えるモデルです。

人体や構造物はより高品質で破綻が少なく、微細なディテールがよりリアルに表現され、自然なテクスチャや影を描写します。

SD1.5ではどのモデルでも生成パラメータは概ね流用可能で、特に気にする必要はありませんでした。

SDXLは一部SD1.5の手法を利用しても問題ありませんが、推奨される生成パラメータがモデルによってもだいぶ変わります。

またLoRAやEmbeddings(EasyNegativeなど)も一切互換性はありませんので、プロンプトの構築も見直す必要があります。

特にSD1.5のネガティブプロンプトでよく使用されているEmbeddingsをそのままXLモデルで入力しても、ただの文字列としてしか認識されていませんので、対応するEmbeddingsに差し替えるか、適切なタグを追加しなければいけません。

このガイドでは、SDXLを使用する際の推奨パラメータ設定について説明します。

2. 基本的なパラメータ

VAE

sdxl-vae-fp16-fix.safetensorsを選択しておけば問題ありません。

モデルに内蔵されている場合も多いですので、指定しなくても大丈夫な場合もあります。

画像サイズ

解像度はTensorArtで用意されているプリセットを使えば問題ありません。

小さかったり大きすぎる解像度は適切な生成結果を得られなくなりますので、SD1.5でよく使用していたサイズはなるべく使用しないでください。

プリセットよりも縦長や横長にしたい場合でも、総ピクセル数を大幅に変更しない範囲で行ってください。

（縦を増やしたら横は減らす等で調整）

例えば1152x896、1216x832、1344x768などがよく使われます。

サンプリング法

モデルによって推奨されるサンプラーがありますので、まずはそれを選択してください。

あとはお好みです。

基本は Euler a か DPM++ 2M SDE Karras あたりを選択しておけば大丈夫です。