No cfg?
https://github.com/Tencent/HunyuanVideo/blob/main/hyvideo/diffusion/pipelines/pipeline_hunyuan_video.py#L1012
The official code here used classifier_free_guidance
But this part of the code is not in the diffusers code，https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video.py#L651
I don't know if I understand this correctly.
I have a question as to how much of a difference it makes whether or not CFG is used in dit
The released HunyuanVideo model is guidance-distilled so it does not make sense to use with true CFG here. While it may still work, it leads to unnecessary 2x more compute and memory requirements, which is why we did not implement the guidance related parts. Even in the official code examples, they only set embedded CFG values and not real CFG, so there is no interpolation between unconditional and conditional latents taking place.