Training Hunyuan Video Checkpoint Finetuning

[aperire]

Hi, are there any finetune scripts or resources for Hunyuan video? It would be a similar format to existing sd-scripts for Flux and SD. Let me know if it is in progress or already exists, thanks!

[Sarania]

Do you mean full fine tuning? At least that I'm aware of, none of the training projects so far support full fine tuning. Both Musubi Tuner(this project) and Diffusion Pipe (https://github.com/tdrussell/diffusion-pipe) support training LORA for HunyuanVideo but not yet full fine tunes. I expect that at some point in the future there will be a way to do so.

Edit: Incorrect, see below!

[aperire]

I see, this also applies for merging lora to checkpoints, right?

[Sarania]

Merging LoRAs to checkpoints can be done in a couple of ways. I've read of people doing it directly in ComfyUI ( https://civitai.com/models/1237378/huncusvid?modelVersionId=1394503 has a workflow example) and also Musubi Tuner comes with a merge_lora.py capable of doing the same. Can read a bit more: #45

[sdbds]

https://github.com/sdbds/musubi-tuner-scripts
This project has supported full-parameters fine-tuning for hunyuanVideo model long ago .
If you are not very familiar with the command line, you can use this script tool.