Question regarding causal=True support in Video VAE Decoder

[xubaixinxbx]

Hi LTX Team,

Thank you for open-sourcing such amazing work!

I have been following recent research on using self-forcing/causal forcing to auto-regressively train a base bidirectional model into a causal model (similar to OmniForcing). This led me to explore the LTX Video Decoder to see if a causal mode is possible.

I noticed the causal=False flag in the current VideoVAE implementation here.

I have a few questions regarding this:

1. Does the current weights natively support setting causal=True?
2. I tested it with causal=True and obtained the results below. Are these artifacts expected behavior under the current weights?
3. If this isn't fully supported yet, are there any plans to release causal decoder weights or provide official support for this mode in the future?

Test Results:
causal=False
https://github.com/user-attachments/assets/85843173-2f8c-4ac5-931e-3f2348cdd679
causal=True
https://github.com/user-attachments/assets/ae50b82c-32c6-40fc-8ed8-04a516a20219

Thanks again for your time and support!