StableVideoDiffusionPipeline cannot use from_single_file

[Meatfucker]

Describe the bug

The StableVideoDiffusionPipeline cannot load models in any format other than diffusers, which is problematic as the latest StableVideoDiffusion model has only been released in safetensors. See: https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt-1-1

Reproduction

svd_pipe = StableVideoDiffusionPipeline.from_single_file("svdxt.safetensors", torch_dtype=torch.float16, variant="fp16")

Logs

Traceback (most recent call last):
  File "D:\ML\apps\svdeez\svdeez.py", line 170, in <module>
    svdeez = SvdeezGUI()
  File "D:\ML\apps\svdeez\svdeez.py", line 16, in __init__
    self.svd_pipe = StableVideoDiffusionPipeline.from_single_file("svdxt.safetensors", torch_dtype=torch.float16, variant="fp16")
AttributeError: type object 'StableVideoDiffusionPipeline' has no attribute 'from_single_file'

System Info

• diffusers version: 0.26.1
• Platform: Windows-10-10.0.19045-SP0
• Python version: 3.10.13
• PyTorch version (GPU?): 2.1.1 (True)
• Huggingface_hub version: 0.20.3
• Transformers version: 4.35.2
• Accelerate version: 0.25.0
• xFormers version: not installed
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help?

No response

[a-r-r-o-w]

The method described in this comment might work if you want to perform the weights conversion to diffusers-expected format.

However, there are a few new checkpoints that have been released and having an improved version of the conversion script as well as .from_single_file mixin implementation would be ideal. @patil-suraj @DN6

[Meatfucker]

Thanks Ill give that script a shot. Was hoping to avoid manually converting things myself but I really want to try the model so appreciate it. Also thank you for tagging the appropriate people as I wasnt quite sure who.

[Yanting-K]

You can use this file to transform this model

https://github.com/Yanting-K/diffusers/blob/main/scripts/convert_svd_to_diffusers.py

[tin2tin]

Getting StableVideoDiffusionPipeline.from_single_file surely wouldn't hurt, as this seems to be the only way to import the new SVD 1.1 release with Diffusers without a conversion. Ex. from here: https://huggingface.co/vdo/stable-video-diffusion-img2vid-fp16/tree/main

[a-r-r-o-w]

Getting StableVideoDiffusionPipeline.from_single_file surly wouldn't hurt, as this seems to be the only way to import the new SVD 1.1 release with Diffusers without a conversion.

Yes, there's a few checkpoints that are out now. I'm working on the single file model import in #6844.

[DN6]

Diffusers format checkpoints have been merged in the stability model repo. https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt-1-1

The model can be loaded using from_pretrained

[tin2tin]

There is one more hindrance, so it can't be downloaded through Diffusers: Repo model stabilityai/stable-video-diffusion-img2vid-xt-1-1 is gated. You must be authenticated to access it.

[Meatfucker]

You can download it via diffusers after authing on the model card page.

[tin2tin]

This works: https://huggingface.co/vdo/stable-video-diffusion-img2vid-xt-1-1

[yiyixuxu]

we can close this one now the diffusers checkpoint has been added?

[Meatfucker]

I think its likely to come up again the next time a new video model gets released. It would probably be ideal to make from_single_file work in the long term.

[CallMeFrozenBanana]

you may try to enter： $ huggingface-cli login and set the token from huggingface

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.