Hunyuan Video LoRA Trainer (fal)

Training a video LoRA usually means renting or owning a serious GPU and babysitting a training run. HunyuanVideoLoraTrainer does it on fal's hardware instead - you feed it a folder of images from inside ComfyUI, it trains a Hunyuan Video LoRA on their servers, and it hands you back a URL to the finished .safetensors. No local training rig, no CUDA fights.

Hunyuan Video (Tencent) is the reason this makes sense as a first LoRA target: it was celebrated early on as easy to train and uncensored out of the box. Its day as the local video model is over - Wan took that crown in 2025 - but for cheap, uncomplicated LoRA training it's still a reasonable place to start. The node is part of gokayfem's ComfyUI-fal-API pack.

How it works

You give it a training set and a step count; fal runs the training job; you get a hosted LoRA file back as a string URL. Because the actual training happens remotely, your machine just uploads images and waits. The pack even includes utility nodes (Load Image Folder, Batch Caption, Images → Training ZIP) so you can assemble the whole dataset without leaving ComfyUI.

Two bits of LoRA-training reality that don't change just because it's on an API: dataset curation beats every knob, and a rare trigger word (something like hy4nvid, not a real word) keeps your concept from colliding with the model's existing vocabulary.

The inputs and outputs that matter

• images (IMAGE) - your dataset, the single most important input. Quality and consistency here decide the result more than any number below.
• steps (100–10000, default 1000) - how long it trains. 1000 is a sane starting point; push higher for a stubborn concept, but watch for overfitting.
• trigger_word - the token you'll type to summon your concept. Use a rare one.
• do_caption (default on) - auto-captions your images before training. Leave it on unless you've captioned by hand.

There's also learning_rate (default 1e-4 - leave it), and two escape hatches for pre-prepared data: images_zip_url (point at a ZIP you already built) and data_archive_format. The single output is lora_file_url (STRING) - a URL to your trained LoRA. Save it or download it; treat it like any other artifact you don't want to lose.

How to install it

Via ComfyUI Manager: search ComfyUI-fal-API, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
pip install -r ComfyUI-fal-API/requirements.txt

Nothing to download locally. Add a fal API key to config.ini under [API] (from config.ini.example) or export FAL_KEY=your_key, restart, and look under FAL/Training.

Common issues & troubleshooting

• Auth error. No FAL_KEY. Set it, restart.
• Training costs real money and takes real time. This is a full training job on fal's GPUs, billed accordingly - it's the priciest kind of call in the pack. Don't kick off a 10,000-step run to test your dataset; do a short one first.
• Weak or overcooked results. Almost always the dataset or the step count, not the learning rate. Trim bad images before you raise steps, and remember the last checkpoint isn't always the best - a shorter run can win.
• The output URL expires. fal result URLs don't live forever. Download the .safetensors promptly, or route the URL through the pack's Save-from-URL utility.
• Windows Portable missing fal-client. python_embeded\python.exe -m pip install fal-client.

Honest framing: you're training a LoRA for a model most people no longer run day-to-day. If your endgame is Wan, train there. But as a no-hardware way to learn video-LoRA training end to end - and for Hunyuan's uncensored, easy-training reputation specifically - this node is a clean on-ramp.