This model has previously been named AnimagineXL Realistic or Stylistic and Make Art Not Goon.
What makes this model special? I prioritized prompt following and composition, making it excellent for art. This makes it excellent at NSFW as a consequence, not by over training on pr0n at the cost of general knowledge, like many other models.
How did I do this? I took the under-appreciated Animagine XL v4.0 because it was well trained from scratch on SDXL 1.0 Base for character knowledge up to Jan 2025. The prompt knowledge of the clip + model is something really special. With the right tags, it is rather artistic and magical with lighting, shadows and composition. CagliostroLab have a discord too. I didn't use "Opt" or "Zero" in the mix, as I think they broke the magic with those.
I think the style and prompt following peaked in v5, and I pushed too hard to improve realism in v6, v7, v8. I also focused too much on integrating DMD weights to get faster generations with all samplers - both normal samplers and LCM samplers.
v9_wip update
It's a work in progress, using the animagine clip with illustrious weights, but maintaining prompt following. For general photorealistic generation, end the prompt with explicit or safe followed by:
sharp focus, depth of field, fine hair strands texture, photo \(medium\), photorealistic, realistic, masterpiece, high score, great score, absurdres
And negative:
ugly, blurry, lowres, worst quality, low score, bad score, text, sketch, drawing,
Previous versions
Check out the workflow in the "Training Data" download! The v8 model is basically tuned for ddpm at cfg 1, 20 steps. However, Euler a works almost as well, and LCM is particularly good as a first pass for architecture, strong shapes, followed by a ddpm pass to add more detail.
Many of the early versions are designed for low CFG. If you want to use cfg > 1 then add the cfg norm node in comfyui to stop burn!
For characters, the prompt format is:
1girl, name, series, rating, s