search

omerbt / TokenFlow

Official Pytorch Implementation for "TokenFlow: Consistent Diffusion Features for Consistent Video Editing" presenting "TokenFlow" (ICLR 2024)
https://diffusion-tokenflow.github.io
MIT License
1.52k stars 134 forks source link

NAN outputs resulting in a blank video #14

Closed rakesh-reddy95 closed 10 months ago

rakesh-reddy95 commented 10 months ago

@omerbt I have run tokenflow using pnp but the results are blank and the decoded latents are NaN. While the inversion during the preprocess happened correctly.

pansanity666 commented 8 months ago

Same here. Any idea about it?

rakesh-reddy95 commented 7 months ago

Which version of SD are you using?

pansanity666 commented 7 months ago

Which version of SD are you using?

I was using SD1.5. I find that several frames get NAN outputs after DDIM Inversion and reconstruction, therefore leads to the following NAN outputs in editing stage. Not sure the reason for it.

rakesh-reddy95 commented 7 months ago

I started using SD 2.1 base version which didn't produce NaN in DDIM inversion and reconstruction. Also, you can try calculating the next step velocity along with the previous step velocity which was removing NaN.