img2img added + update input descriptions #23
Conversation
|
All the code is from here https://github.com/Gothos/diffusers/tree/flux-inpaint/src/diffusers/pipelines/flux Gothos actually works at fal this is the img2img source code https://github.com/Gothos/diffusers/blob/flux-inpaint/src/diffusers/pipelines/flux/pipeline_flux_img2img.py |
predict.py
Outdated
| ) | ||
| # Override width and height | ||
| width, height = resized_width, resized_height | ||
| flux_kwargs["image"] = input_image.resize((width, height), Image.LANCZOS) |
There was a problem hiding this comment.
Resizing creates artifacts, could we crop to the nearest smaller multiple of 16 instead?
- Crop images for img2img and inpaint modes (from center) - Make dimensions multiples of 16 (round down) - Avoid resizing artifacts
|
@zsxkib |
|
Hey @mountainguan, thanks for bringing this up. What you're seeing is likely due to our safety checker kicking in. You can actually bypass this when using the API by setting Just a heads up though - this safety feature can't be turned off on the website version. If you're curious about how we handle platform safety, check out our docs: Replicate's approach to safety Let me know if you need any more details! |
TLDR: Added image-to-image (img2img) functionality.
• Created img2img pipelines for dev and schnell models
• Integrated img2img into the predict method
• Added logic to detect img2img mode from inputs
• Improved handling of input images and masks for inpainting
• Adjusted image resizing for inpainting
• Updated input descriptions for width, height, and aspect ratio
• Added explanations for when inputs are ignored
• Allowed input images to override aspect ratio and dimensions
• Kept compatibility with existing txt2img feature
• Reused pipeline components to reduce memory use