replicate / cog-sdxl

Stable Diffusion XL training and inference as a cog model
https://replicate.com/stability-ai/sdxl
Apache License 2.0
212 stars 106 forks source link

"refine_steps" argument name and docs are misleading #54

Open iuliaturc opened 5 months ago

iuliaturc commented 5 months ago

When setting refiner=base_image_refiner, the refine_steps argument promises to control how many denoising steps will be performed with the refiner:

Screenshot 2024-05-09 at 11 39 35 AM

However, in practice, this is only true if you additional set prompt_strength=1.0. The reason is that:

This results in surprising behavior, where you set refiner_steps=N only to see that it ran for fewer steps. I think you could either: