32 karma · joined December 15, 2023
Hopefully we can animate your bear cartoon one day!
We are trying to better understand the model behavior at the very end of the video. We currently extend the audio a bit to mitigate other end-of-video artifacts (https://news.ycombinator.com/item?id=41468520), but this can sometimes cause uncanny behavior similar to what you are seeing.
We use rectified flow for denoising, which is a (relatively) recent advancement in diffusion models that allow them to run a lot faster. We also use a 3D VAE that compresses the video along both spatial and temporal dimensions. Temporal compression also improves speed.
We're also very excited about the template idea! Would love to add that soon.
Edit: If we generate videos at a lower resolution and with a fewer number of diffusion steps compared to what's used in the public configuration, we are able to generate videos at 20-23 fps, which is just about real-time. Here is an example: https://6ammc3n5zzf5ljnz.public.blob.vercel-storage.com/fast...
We actually did this in early overfitting experiments (to confirm our code worked!), and it worked surprisingly well. This is exciting to us, because it means we can have actor-specific models that learn the idiosyncratic gestures of particular person.