Inversion Results

[abcdvzz]

Hi I only tried the edm inversion_video code, the reconstruction results are bad.

FYI, I used the code here: https://huggingface.co/docs/diffusers/en/using-diffusers/svd, and set param latents = inv_latents, I also changed the following code to just use image_latents and image_embeddings in the unet:

# image_latents = torch.cat(
#     [
#      image_latents.unsqueeze(1).repeat(1, skip, 1, 1, 1),
#      firstframe_latents.unsqueeze(1).repeat(1, latents.shape[1]-skip, 1, 1, 1)
#     ],
#     dim=1
# )
image_latents = image_latents.unsqueeze(1).repeat(1, latents.shape[1], 1, 1, 1)

Is there anything else I should notice? What about your results with only edm inversion?

[Vicky0522]

You can directly try EDM inversion using this repository by disabling other parameters and enabling inversion (use_inversed_latents=true) in the configuration files. Hope this will be helpful for you.