PixArt-Sigma-1024px_512px-animetune

If you would like to know the prompts for the sample images or detailed explanations of the models, please refer to the information on civitai.

https://civitai.com/models/505948?modelVersionId=562359

I don't think shakker yet supports Pixart image generation.

A good setting for cfg is between 2-5.

For inference, I like SDE because it gives a similar feel to the dataset, so I generate it with 12-15 steps.

Euler A is stable with few breakdowns. It's blurry, so a higher cfg or step is better.

■Here are my recent favorite inference settings. This will be updated as needed.

This is not the optimal solution.Please try various things!

Both booru tags and natural language are available for use.

Using the custom node "Uncond-Zero-for-ComfyUI".

sampler:"SDE cfg4 step12" ,"Euler cfg_pp" or "Euler A cfg_pp" cfg1-2 step40-50

Scheduler:"GITS" or "simple"

Using Uncond-Zero-for-ComfyUI for automatic CFG adjustment: GITS provides rich textures, Simple ensures stable generation quality, SDE stays true to the dataset, and Euler offers stability.

SDE performs well with a low number of steps, but anatomical accuracy improves with higher steps like 50. Once you find a good seed, increasing the number of steps is recommended.

Negative prompts are not trained. Please try various prompts!

As described in the dataset contents on the page below, if you don't like realistic textures, you might want to include terms like "realistic, figure".

I don't like restrictions and prioritize diversity, so I keep the negative prompts to a minimum.

However, my way of thinking is unconventional. You don't have to follow it! You might get better results with many negative prompts, so give it a try!

■This is an experimental fine tuning.

Attention This Fine Tuning model is very difficult!

The quality is not good!! Don't expect too much!

I would be happy if you could be interested in Pixart even a little.Pixart has potential.

■I trained using onetrainer.

Fine-tuning is performed on a 70,000-image dataset that mainly contains anime images, but also some realistic and AI images. The training resolution is 512px or 1024px. Pixart is high quality but has low requirements, making it suitable for training.

■1024px_v0.1 has been updated.

Tag training has become comparable to the 512px model.

Anatomy has improved compared to before. There are still many flaws, though...

However, I recently realized that instead of worrying about and fixing flaws, I find it more enjoyable to come up with prompts that create images with a good atmosphere. It's enjoyable to provide ChatGPT with a theme and have it generate natural language prompts!

Generating with natural language is fun, but the accuracy might have decreased due to the progress in tag training. Therefore, it might be interesting to merge it with the previous model or other models fine-tuned with natural language.

For those who don’t want to waste time, I think it’s a good idea to practice with the 512px model first to see which tags work before using the 1024px model.

■pixart_1024_13_p2_ep12_merge

add Merge experimental model. slightly more realistic style due to averaging, and the balance might be better. But there's also a half-baked possibility.I quietly added it; this model is very experimental, so it is a secret between you and me who found it...