arkennemasis fal · MiniMax Hailuo 2.3 Pro Text to Video · $0.49/video

Text in, a 1080p clip out, $0.49 flat. That's the whole interface. Which is either exactly what you need - an establishing shot, a bit of texture, a clip to drop behind narration - or a trap, because with no first frame you've handed the model every decision about what this clip looks like.

When text-to-video is the right node

If you already have a still, use the image-to-video version of this node instead. Same model, same price, and you keep control of the composition. Text-to-video earns its keep when there's no still to start from: clouds over a skyline, hands working at a bench, an abstract transition, a scene-setting wide you'll cut away from in two seconds.

The case against it is consistency. The community's character-consistency answer has moved firmly toward "generate once, then edit or animate that image" precisely because re-rolling a look from text gives you a different look each time. If a person or a product has to match across five shots, generate the frame in your own pipeline and animate it with the i2v node. Text-to-video is for shots nothing else has to match.

How it works

Under the hood it's the pack's standard fal wrapper around minimax/hailuo-2.3/pro/text-to-video: pull FAL_KEY from the .env next to your launcher, validate the inputs, estimate the cost and compare it to max_cost_usd, submit one billed request, poll it while the node reports queued / running / seconds, then download into output/fal/fal-ai_minimax_hailuo-2.3_pro_text-to-video/. Nothing before the submit costs anything, so a rejected input is free. Hailuo itself is MiniMax's closed, hosted video line - nothing here runs on your GPU, and there's no local version to fall back on.

Inputs

• prompt - the entire steering wheel, capped at 2000 characters by the schema. Name the subject, the action and the camera move, in that order; leave the model room on lighting and texture unless you actually care. Over-specifying a ten-second action across a long paragraph is a classic way to get mush.
• prompt_optimizer (default on) - the model rewrites your prompt before generating. Useful when you've written three words and walked away; turn it off when the wording is deliberate, because it will be reworded.

Four optional controls, and two of them matter in practice:

• max_cost_usd - default $20, 0 disables. The node refuses before uploading when the estimate is over the cap.
• max_concurrent - default 1, so a canvas stacked with fal nodes calls fal one paid job at a time instead of twenty at once.
• reuse_identical_run - default on. Identical inputs and a still-present result file means you get that file back for free rather than a second bill. Turn it off when you want a genuinely new roll of the same prompt.
• extra_json - raw JSON merged into fal's request last, for any field the node has no box for.

Outputs are video (a real ComfyUI VIDEO - feed Video Save or the pack's Video Assemble), video_path with the on-disk file, and info, a JSON blob holding the request id, the estimate, the saved files and fal's full response. No audio: this endpoint returns picture only. If you need the model to invent sound as well, that's a different family in the same pack (the MiniMax H3 nodes), not something to expect here.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install httpx

Or ComfyUI Manager → Install via Git URL, pasting the repo URL. Then create the .env next to run_nvidia_gpu.bat:

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

Restart. It lives under arkennemasis/fal/Video/MiniMax, and the menu title carries its live price.

Traps

The usual two thirds of fal-node pain is one thing: the key. FAL_KEY must be in the .env beside the launcher (or in the ComfyUI folder) - there is no key box on the node, on purpose, so a missing key is a silent-looking nothing until you read the error. Second: at $0.49 a roll, "let me just try six prompts" is three dollars before lunch. Draft your prompt wording on the 768p Standard text-to-video node at roughly half the price, then render the take you like here.

And if a run vanishes after a ComfyUI restart, it didn't fail - it finished on fal while you weren't watching. The request id is in output/fal/_requests.jsonl; fal Recover Result pulls the finished job down for free.