Our latest 8Buff SDXL Model 8Buff_XL_v2 has been released!

Model notes and prompt suggestions:

• Based on our 8Buff_XL_V1 model, we have fine tuned, mixed, and trained our model to improve the image quality, diversity, and consistency.

• Our models tend to focus on the generation of multiple people. Use terms Woman/Man for singular, and Men/Women for multiple. You can provide a number (3Men) or specify a line of people (line of cheerleaders) also. Stating things like ‘people in background’ will provide people and crowds, but will often be blurry and less defined.

• To generate different skin colors and looks, you can use countries names or geographic regions such as; Nigeria, European, Canadian, Japanese, ect…

• Use terms like Man/Woman for adults, girl/boy for younger people.

• When creating groups of people that should look different from one another use positive prompt terms such as; different face, different hairstyle, different body type, etc. To enforce this you can also use negative prompts; same face, same clothes, etc…

• Training done primarily on 1024x1024, 1024x1280, and 1280x1024 images. When mixing with LORA(s) recommend starting with 1024x1024 and adjusting from there.

• Sampling method that provides the most consistent results is DPM++ 3M SDE Karras (in Automatic1111). Any of the Exponential methods tend to generate more refined images but less consistency.

• Known issues:

  • Some sample methods tend to paint color streaks over pictures.

  • Perspective, hands, and eyes are still issues, like most models. Using seed and variation adjustments can help address some of these.

  • It is biased and tends to be a little thirsty at times. Be selective with your prompts.

  • Still very bad at spelling, but you can do basic text with prompts like, small sign with handwritten text ‘eightbuff’, with the text you want between the ‘ quote' signs.

Works well with LORAs.

Check out our blog https://eightbuff.blogspot.com/ for more details.