Question about finetuning with different resolution and frame nums

[zqh0253]

Thank you for providing such an excellent codebase. I'm curious if this repository supports fine-tuning the model (e.g., CogVideox) with different resolutions and frame numbers compared to the default setting, such as 256 x 256 x 6 (H x W x T). If so, what is the best practice for doing this?

[neph1]

I've done some experimentation with this with LTX-V. Multiple resolutions add flexibility to the lora, ie results get better when using different resolutions during inference (which is expected, I guess).
LTX doesn't seem to like lower framerates. I've tried to train it at as low as 12 fps, but seem to be getting smeared results as it interpolates to 24 fps. Although it's possible to offset it somewhat by raising the fps during inference (40fps, 60fps or more). I'd be happy to be corrected if anyone has better results.