arkennemasis fal · MiniMax Hailuo 2.3 Fast Pro Image to Video · $0.33/video

You have a still you're happy with and you want it to move, at 1080p, without babysitting a 14B video model's VRAM. That's this node's entire job. It wraps fal's hailuo-2.3-fast/pro/image-to-video endpoint - MiniMax's hosted Hailuo line, the one behind the Hailuo video service - and it charges a flat $0.33 per video.

Why you'd reach for it instead of a local model

Hailuo has no downloadable weights. MiniMax is API-first: the KB's picture is a Shanghai lab that built Hailuo into one of the closed video names people point at next to Kling and Veo, and only opened anything in August 2026 with H3 - whose community licence geofences the US, EU, UK and Korea out of local use anyway. So for a lot of people this isn't "cloud is convenient," it's the only door.

The practical reason is the price ladder inside this pack. Fast Standard at 768p is $0.19 for 6 seconds and $0.32 for 10; plain Standard at the same resolution is $0.28 and $0.56. This one is a flat $0.33 for 1080p Fast, and straight 2.3 Pro is $0.49. So Fast Pro is the 1080p tier that doesn't cost Pro money - and if 768p is good enough for your cut, it's the more expensive of your two options at both lengths. Decide resolution first, then price.

How it works

Not a model on your machine. The node reads your FAL_KEY, checks the inputs, estimates the cost, uploads your first frame to fal's CDN (free), submits one billed request, polls it while showing status, then downloads into output/fal/fal-ai_minimax_hailuo-2.3-fast_pro_image-to-video/. Everything up to and including the upload costs nothing - a failure there never bills you.

The inputs that matter

There are only three required ones, which tells you how little fal exposes here:

• prompt - capped at 2000 characters in the schema. Describe the motion and the camera, not the picture; the picture is already the first frame. One action per clip lands far more reliably than a paragraph of choreography.
• image - your first frame. The node checks for a batch where a single picture is expected, so feed it one image, not a batch, or you'll get told off before anything is billed.
• prompt_optimizer (default on) - the model rewrites your prompt before generating. Handy for a loose sentence, less handy when you've tuned exact wording; turn it off if you want your text used verbatim.

Notice what isn't there: no duration box. fal bills this endpoint per video rather than per second, and the node mirrors fal's input list exactly, so there's nothing to tune. Want the 6-or-10-second choice? That lives on the Standard 2.3 image-to-video nodes.

Optional: max_cost_usd (default $20 - the node stops before uploading if the estimate exceeds it; 0 disables the cap), max_concurrent (default 1, so a canvas full of fal nodes goes one paid call at a time), reuse_identical_run (default on - same settings and same picture returns the already-paid-for file instead of billing again), and extra_json for any fal field with no box.

Outputs: video (a real ComfyUI VIDEO) plus video_path with the file on disk, and info, a JSON blob with the request id, the cost estimate, the saved files and fal's full answer. Wire video into Video Save or the pack's Video Assemble; take the request id from info if you ever need fal Recover Result. There's no audio track - this endpoint returns video only.

Install

In ComfyUI Manager search the pack title, or paste the repo URL into Install via Git URL. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install httpx   # portable: python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\comfyui-arkennemasis\requirements.txt

Then put your key in the .env next to run_nvidia_gpu.bat and restart:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

There is deliberately no key box on the node - a typed key ends up in every workflow you share.

Where people get burned

FAL_KEY not found is the number one failure, and it's almost always the .env sitting in the wrong folder - it belongs beside the launcher, or in the ComfyUI folder, never in the custom node's directory. Second: a restart mid-wait. The job keeps running on fal and you keep paying, so grab the id from output/fal/_requests.jsonl and pull it down free with fal Recover Result. Third: a wall of fal nodes uploading at once; max_concurrent of 1 exists because twenty nodes hitting fal in the same second is how you find out what a 429 looks like. And the one nobody remembers: reuse_identical_run handing back a previous take and looking like the node ignored your click. It didn't - it saved you $0.33.