[lcolucci]

Our transformer model was trained to generate videos that are up to 8s in length. However, we can make videos that are longer by using it an an autoregressive manner, and taking the last N frames of output i to seed output (i+1). It is important to use more than just 1 frame. Otherwise ,the direction of movement can suddenly change, which looks very uncanny. Admittedly, the autoregressive approach tends to accumulate errors with each generation.

It is also possible to fine-tine the model so that single generations (one forward pass of the model) are longer than 8s, and we plan to do this. In practice, it just means our batch sizes have to be smaller when training.

Right now, we've limited the public tool to only allow videos up to 30s in length, if that is what you were asking.

[leobg]

Video compression algorithms use key frames. So can’t you do the same thing? Essentially, generate five seconds. Then pull out the last frame. Use some other AI model to enhance it (upscale, consistency with the original character, etc.). Then use that as the input for the next five seconds?

[andrew-w]

This is a good idea. We have discussed incorporating an additional "identity" signal to the conditioning, but simply enforcing consistency with the original character as a post-processing step would be a lot easier to try. Are there any tools you know of that do that?

[naveensky]

Thanks for answering this. I would love to use it when APIs are available to integrate with my apps