Inquiry about video extension

[bcb225]

Hi, thanks for the great work.

I have a quick clarification question about video extension.

Does the current open-source Causal Forcing code officially support short-video continuation / video extension with a multi-frame input video prefix, beyond I2V?
For example, if I encode a 3-second input video into latents and pass it as initial_latent, is this considered a supported use case for generating the following future frames, or is it still more of an internal / experimental path?

I saw that initial_latent with more than one frame is described as video extension in the pipeline code, but I could not find a clear user-facing example in the main inference script.

Thanks!

[zhuhz22]

Hi @bcb225 ,
Yes it supports. You can set the initial_latent the input video. But note that the total length of video (input+generated) should be 5s, if you use the default ckpt. For longer video, refer to our long video generation section.
If this is a common command, we will consider officially add the procedure into the REDME. Currently, you can try it yourself, and we are willing to know your results.
If you have succeeded, we sincerely welcome you raise a PR about this.❤️

[bcb225]

I have one more question specifically about 1:1 / square video resolution.

Does the current open-source Causal Forcing code support square inputs/outputs?

I am asking because, from reading the released inference code, it looks like the current setup may be tied to the default Wan latent shape. For example, the public inference scripts seem to use fixed noise shapes such as:

• inference.py: [1, args.num_output_frames, 16, 60, 104]
• long_video/inference.py: [args.num_samples, args.num_output_frames, 16, 60, 104]

So I am not sure whether square resolution is already supported in the current codebase, or whether it would require modifying the latent/noise shapes and related inference pipeline logic.

Also, if I understand correctly, Wan itself seems more flexible with resolution.

If square resolution is not currently supported, would it be possible to enable it with relatively small code changes to the released inference pipeline, or would it require additional training / fine-tuning to work properly?

Thanks!

[zhuhz22]

We have not tested it, and you can simply modify the kv cache size to support it. We only trained the model at 480$\times$832, so it may suffer quality degradation, but we are not sure how severe the degradation will be.

The hard-coded inference are inherited from the Self Forcing code. We suggest trying to infer in a training-free way first, and if the degradation is very severe, additional training / fine-tuning is needed.