How to build a dataset for finetuning CogVideoX I2V 1.5

[royvelich]

Hi,
I want to finetune the CogVideoX I2V 1.5 (5B) model. I have a set of videos that I want to use, but first I need to preprocess them so they meet the requirements of the model. Do I have to make sure that my fine-tuning dataset meets the generation properties of the model? That is, in the case of CogVideoX 1.5, the videos should be:

• Min(W, H) = 768
• 768 ≤ Max(W, H) ≤ 1360
• Max(W, H) % 16 = 0
• Video Length: 5 seconds or 10 seconds
• Frame Rate: 16 frames / second

Do I need to make sure that all my fine-tuning videos follow those guidelines?

[sayakpaul]

These are recommended settings to ensure the results don't deviate because of discrepancies. But I think we can still fine-tune it hard enough to adapt to particular settings. Some models are easier to adapt than others (LTX, for example, is quite adaptable).