Is there an existing issue for this?
Contact Details
No response
What should this feature add?
This FR is a companion to #8763. HiDiffusion is a training-free method for higher-resolution image generation with SD1.x and SDXL models. It would be particularly useful for allowing SD1.x models to generate and inpaint at standard SDXL resolutions without a separate high-res fix.
The most up-to-date code for HiDiffusion is located with dgenerate here. The author's notes for these changes are here.
Alternatives
No response
Additional Content
cc @Pfannkuchensack
Is there an existing issue for this?
Contact Details
No response
What should this feature add?
This FR is a companion to #8763. HiDiffusion is a training-free method for higher-resolution image generation with SD1.x and SDXL models. It would be particularly useful for allowing SD1.x models to generate and inpaint at standard SDXL resolutions without a separate high-res fix.
The most up-to-date code for HiDiffusion is located with dgenerate here. The author's notes for these changes are here.
Alternatives
No response
Additional Content
cc @Pfannkuchensack