diffusers icon indicating copy to clipboard operation
diffusers copied to clipboard

Published 20 hours ago • verified publisher icon huggingface

Stable-diffusion Image-to-image pipeline is broken for batch_size > 1

Open jorgemcgomes opened this issue 3 years ago • 1 comments

Describe the bug

When multiple images are passed to the init_image argument as a tensor, init_latents will have the same dimension 0 as init_images. But this line is further duplicating the init_latents by the batch size, which results in a mismatched tensor size.

https://github.com/huggingface/diffusers/blob/f3983d16eed57e46742d217363d8913bef7f748d/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py#L288

The fix is trivial: init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0)

Reproduction

No response

Logs

No response

System Info

diffusers-0.5.0-dev (master)

jorgemcgomes avatar Oct 07 '22 19:10 jorgemcgomes

+1

Thanks for the issue! Will open a PR now

patrickvonplaten avatar Oct 10 '22 13:10 patrickvonplaten