HunyuanVideo 1.5 LoRA training: train.py and the Muon optimizer

Yes, HunyuanVideo 1.5 has official LoRA training code. The README's Dec 05, 2025 news line says training code and a LoRA tuning script were released with Muon optimizer support, and the README states that to continue training the model or fine-tune it with LoRA you should use the Muon optimizer. The entry point is train.py, with --use_lora, --lora_r 8 and --lora_alpha 16 in the example.

What does the training setup need?

From the README: distributed training launched with torchrun, FSDP support, and gradient checkpointing. It names Muon as the optimizer to use for continued training and LoRA. The README gives no VRAM figure for training, and I did not find a dataset-size or training-time example, so budget by testing a short run first.

What does a hosted API offer instead?

Not a LoRA. In v1, Sume's allowed_passthrough_parameters list is empty for every video model and a non-empty provider.options returns 400 unsupported_parameter, so there is no field to attach your own weights to. The route for a look or a character on Sume is reference media: input_references for style or content, or frame_images for a first or last frame, on the models that list them.

So the choice is concrete. If your product depends on a trained style that must be applied to every clip, you need open weights and your own GPUs, and the license that comes with them. If a few reference images are enough, a hosted id is less work. Sume does not list HunyuanVideo (catalog code, read 2026-10-09); see the video docs for what is listed.

What should a first training run look like?

Start small: a short clip set of one subject or one style, a low rank as in the README example (--lora_r 8), and a few checkpoints saved early. Test each checkpoint on prompts that are not in the training set, to see whether you have learned the look or memorized the clips. The README does not give recommended dataset sizes or epochs.

Before you train on anyone's footage, confirm you have the right to use it. Training data rights are your responsibility, and check the base weights' license for what it says about derivatives.