This model has been trained on 12k images based on the Pixart-Sigma model (it is not compatible with Stable Diffusion!). To use this model, please refer to this article :

https://civarchive.com/articles/5267/pix-art-sigma-sdxl-workflow-15-15-steps

The file you download should be placed in the checkpoints folder. So this model is very NSFW, but it follows the prompt very well because it is a Pixart-Sigma model.

You need to use fully-formed sentences and not WD14 tags to fully use this model's potential. A nice trick is to ask chatgpt to generate the prompts for you.

To control the nudity, you can try to put this in the negative prompt:

1. adult move model

2. she has her beasts exposed

3. she has her vagina exposed

4. she is totally naked

5. she is partially naked

These mostly suffice to eliminate the nudity. When it is not, there is most likely something in the positive prompt that triggers it.

In the initial release of this model, a 512 pixels multi-step model is released. The ideal resolutions for it are:

• 512x512

• 512x768

• 768x512

In the upcoming version, the ideal ones are:

• 1024x1024

• 800x1200

• 1200x800

Also note it is only usable through comfyui as there is no extension for it in Automatic1111 at the time of writing. Since Pixart-Sigma is new compared to Stable Diffusion, some of the tools in comfyui will not work. Those tools are:

• Face detailer

• Iterative upscale

• Latent upscale

Ultimate Upscale do work with Pixart-Sigma. I highly suggest using Face Detailer with a SD 1.5 model to get good results with the 512 pixels model.

PS . It includes watermarks, it requires some additional training data to get rid of those...

The primary dataset source is from Score Classic, scraped from pornpics.com/