MiniMax Subject Reference (fal)

Want the same face - the same person - to show up in a generated video looking like themselves? That's what MiniMaxSubjectReference does. You hand it one reference image of a subject and a prompt, and MiniMax (Hailuo) generates video that keeps that subject consistent. It's the character-consistency trick, done through a single reference image, no training required.

It runs on fal's servers via gokayfem's ComfyUI-fal-API pack - no local model, billed per clip. Reach for it when you have a specific character or person and you want them, not a random stranger, in the shot.

How it works

You provide a subject_reference_image and a prompt; MiniMax uses the reference to anchor the subject's identity while generating a new scene from the prompt. fal runs it remotely and returns a URL to the clip. The output is a string URL, so save it before the link expires.

The inputs and outputs that matter

• prompt (required) - the scene and action. Describe where your subject is and what they're doing.
• subject_reference_image (required, IMAGE) - the identity anchor. A clean, clear, well-lit shot of the subject's face works best; a busy or low-res reference gives the model less to hold onto.
• prompt_optimizer (default on) - lets MiniMax rewrite your prompt for better results. Handy for short prompts; turn it off if you want your wording respected exactly.

The single output is a STRING - the video URL.

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

No models download. Add a fal API key to config.ini under [API] (from config.ini.example) or export FAL_KEY=your_key, restart, and find it under FAL/VideoGeneration.

Common issues & troubleshooting

• Auth error. FAL_KEY not set. Fix and restart.
• The subject doesn't look like the reference. Give it a better anchor - a sharp, front-facing, well-lit image of the subject. One good reference beats a cropped or blurry one, and consistency is never perfect, so expect some variance.
• Your prompt got changed. That's prompt_optimizer. Turn it off to keep your exact wording.
• The URL expired. fal result links don't last - route the STRING output into the pack's Save-Media-from-URL utility or download it promptly.
• A run fails. The pack raises fal's real error message - content policy, quota, or a bad input. Read it. Note that a recognizable-person reference can trip content policies.
• Cost. Each clip is a paid API call. The result cache serves repeat identical calls free (force_rerun bypasses); the spend guard caps a session.
• Windows Portable. Missing fal-client? python_embeded\python.exe -m pip install fal-client.

For a recurring character across multiple clips, this single-reference approach is about as low-effort as consistency gets - no LoRA, no training, just a good headshot and a prompt. If you need several subjects kept consistent at once, the Kling Omni reference node in the same pack handles multiple elements.