Model introduction I Introduction

In this introduction, you'll learn:

1. model introduction (see section II); usage guide (see section III);
2. training parameters (see section IV);
3. trigger word list (see
Appendix A)

II model introduction anime illustration design XL, or

AIDXL, is a special purpose

Model for generating two-dimensional illustrations. It has over 200 built-in illustration styles (and more as updates increase), which are triggered by specific trigger words (see Appendix A)

Advantages: The composition is bold, there is no sense of pose, the main subject is prominent, there are not too many complicated details, and I know many anime characters (inspired by the Japanese spelling of the characters, for example, “ayanami rei” corresponds to the character “Ayanami Rei”, and “kamado nezuko” corresponds to the character “Nezuko”).

The model is difficult and is not recommended for beginners.

III. The User Guide (which will evolve with the times) recommends using ComfyUI

to generate images... Currently, there is no noticeable difference between WebUI and ComfyUI

in terms of generation.

1 Generation parameters If you are unable to generate an image similar

to the preview image, please refer to the following guide

1. It is recommended that the total image resolution (total resolution = height x width) be greater than 1024x1024 and less than 1024x1024x1.5; otherwise, the resulting image may be of poor quality. This is a rule of thumb, that is, the total resolution of the generated images should be higher than the total resolution of the training set images, and at the same time less than 1.5 times the total resolution of the training set images to prevent blurring and distortion. For example, this model was trained at a total resolution of 1024x1024, so you can generate images up to 1024x1536 (2:3 for example) resolution
. It is
2. recommended to write positive reminders in the form of tag + natural language. Increase the density of nouns in natural language and avoid using abstract adjectives or using multiple adjectives to modify nouns on top of each other. Also, there's no need to use too many negative reminders. It is recommended that the number of negative reminders be no more than 10
.
3. The “Clip Skip” operation is not performed, that is, Clip Skip = 1.
4. Using the “dpmpp_2m” sampler (sampler), with the “karras” scheduler (scheduler), this combination is called DPM++ 2M Karras in webui. Sample more than 35 steps on the 7 CFG Scale
.
5. You only need to use the model itself, not the Refiner.
6. Use the base model vae or sdxl-vae.
7. Use the trigger words provided in the appendix section to take advantage of stylization. Note that starting with v0.5, some quality indicators, such as best quality, masterpiece, etc. will be supported. Using them will improve the average aesthetic quality of images (not always)

2 Note

• Use the VAE model, text embedding (embeddings) model, and Lora model supported by SDXL. Note: sd-vae-ft-mse-original is not a SDXL-enabled VAE; negative text embeddings such as easyNegative and badhandv4 are not SDXL-enabled embeddings.
• When generating images, it is highly recommended to use model-specific negative text embedding (see download https://civitai.com/models/144327/negative-embeddings-aidxl-series-models), since it is specially made for the model, it has almost only a positive effect on the model
.
• Due to initial training, the new trigger words added to the version will be relatively weak or unstable in the current version.

3 The styles indicated by the experimental

• trigger words
can blend with each other to create a new style.

IV. Using SDXL1.0 as the base model, model A was obtained by training

for about 100 sessions on a cosine scheduler with a 5e-6 learning rate and a total length of 1 using approximately 2w self-labeled images. After that, with a learning rate of 2e-7 and the same parameters, model B was trained to obtain model B. The AidxLv0.1 model is obtained by mixing models A and B

V vs. SD1.5-based AID 2023/08/08: AIDXL

trains using the exact same training set as AIDv2.10, but performs better than AIDv2.10. AIDXL is smarter and can do many things that the SD1.5 base model can't do. It's also great at distinguishing between different concepts, learning image details, processing compositions that are difficult for SD1.5, and almost perfectly learning styles that older AID couldn't fully master. Overall, it definitely has a higher limit than SD1.5

Appendix A. List of

trigger words I'm sorry.

For security reasons, there is no list of prompts on this page. See Appendix A of the CivitAI model page for the full list of prompts: https://civitai.com/models/124189/anime-illust-diffusion-xl