Possible to get a better VAE?

Open jpgallegoar opened this issue 9 months ago • 4 comments

Dear LTX team, the provided VAE is not very good for the multiple keyframe workflow. The image degradation dure to latent compression on frames other than the first one is destroying all chances of having a good LTX video interpolator, and I believe it's the only hurdle for this task. I have attached two examples where you can see my issue.

https://github.com/user-attachments/assets/46c051fc-1a2b-40be-9706-a4e025c416b2

Mar 07 '25 10:03 jpgallegoar

+1, I hope to get a response to this question

Mar 09 '25 12:03 wangguangyuan

I'v found same problem when I reconstruct video of human, the eye and mouth regions degrade obvious.

Mar 12 '25 06:03 2502128021

this is why i check the github issues before adopting a repo lol

Mar 17 '25 02:03 ksasso1028

Yes, we working on this.

Mar 18 '25 21:03 yoavhacohen