It's a hybrid of "Illustrious" and "SDXL" (although it's more "Illustrious"), so some lora may not work. It only works properly with the "LCM" sampler. Good results can be achieved using: "euler_a/karras 6-8 Steps, lcm/beta 6-10 steps" (applies to turbo version only). The emphasis was on realism. The settings will be in the picture.
Description
More SDXL, more realistic, but less understanding of the prompt. At every opportunity, she tries to undress the girl. Less artistic. Less understanding of concepts. Overall, the model is less stable.
FAQ
Comments (3)
After playing around with v2.5, I agree with your description. It's like 5% more realistic in terms of texture but has lost 20% or more of its prompt adherence. That's my first impression, anyway.
By the way... these models have enough DNA from Illustrious that you can replace the CLIP with something like this and it seems to understand danbooru tags even better without losing much (if any?) photorealism or LoRA compatibility.
Looking forward to more updates!
Yes, I've already downloaded them and am using them, but not fully. I've mixed them with my own CLIP. I'll try using them exclusively, and if that improves the results, the next version will be based on these CLIP.
@toya_san Mixing is probably the way to go. I've run into a few instances where the model only understands certain danbooru tags if I use 100% Illustrious CLIPs, but overall, I think your built-in CLIPs have superior photorealistic texture.
Sprinkling in like 30% IL might help with image composition, though.