[Feat]: HunyuanVideo 1.5 support

[rkfg]

Describe your use-case.

Since HunyuanVideo 1.0 is supported, it would be great to support 1.5 as well. This model is smaller (8B) but can do things better. There's camera control, better prompt following, faster inference and so on. Multiple models were released for T2V, I2V, upscale, and their 480p/720p/distilled variants. I suppose they have the same base so a lora trained on one model would apply to all of them but that's yet to be seen.

Code: https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5
The original models are here: https://huggingface.co/tencent/HunyuanVideo-1.5
Repackaged for ComfyUI here: https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged

It's already natively supported in ComfyUI.

What would you like to see as a solution?

Lora and optionally full model fine tuning. This time the model is available without CFG distillation which should improve the lora quality I think. The distilled variants don't seem to have the embedded guidance. The architecture is different, just loading 1.5 with the existing 1.0 code will not work.

Have you considered alternatives? List them here.

I love the OneTrainer's tools for dataset processing in particular so it's my go to training software. However, diffusion-pipe developer plans to add HyV1.5 support soon.