arkennemasis fal · MiniMax Hailuo 2.3 Standard Text to Video · $0.28-0.56/video

Pure prompt to 768p video: $0.28 for 6 seconds, $0.56 for 10. It's the cheapest way to ask a closed video model "does this idea work at all?" before committing to 1080p - and honestly, that's how you should treat it.

What it is

One node per fal endpoint is the rule in this pack, and this one wraps minimax/hailuo-2.3/standard/text-to-video. Hailuo is MiniMax's closed video line: hosted, no weights, nothing to run locally. That being the case, the sensible way to use a hosted video model is to spend as little as possible while you're still deciding what you want.

The three text-to-video rungs here: Standard 768p at $0.28/$0.56 for 6/10 seconds, Pro 1080p at $0.49 flat, and the two image-to-video families either side. If you have a first frame, the i2v nodes give you more control for the same money. Text-to-video is for the shot with nothing to anchor it - the wide, the establishing beat, the texture clip behind narration.

How it works

Read FAL_KEY from the .env beside your launcher, validate, estimate the cost, submit one billed request to fal, poll it, download the result to output/fal/fal-ai_minimax_hailuo-2.3_standard_text-to-video/. Every step before the submit is free, including the uploads when there are any - here there are none, since text-to-video sends no media. That means a run that fails input validation costs you nothing at all.

Worth knowing: because fal's own billing here is per second of generated video, the node's duration widget is the price control, and the live badge above the node updates as you move it.

Inputs

Three, and only one is really yours:

• prompt - up to 2000 characters. Subject, action, camera. Written like a shot note, not a novel.
• duration - 6 or 10. $0.28 or $0.56.
• prompt_optimizer (default on) - the model's own prompt rewriter. Keep it on while you're exploring; turn it off when you've arrived at wording you want used as written, because it will otherwise be paraphrased for you.

Optional: max_cost_usd ($20 default; stops the run before anything is sent when the estimate exceeds it, 0 turns the cap off), max_concurrent (1 by default - one paid fal call at a time in the whole run), reuse_identical_run (default on - identical settings return the file you already paid for, restart included), extra_json (JSON merged into fal's request last, for fields the node doesn't expose; {"seed": 7} is the canonical example).

Outputs: video (ComfyUI's VIDEO, ready for Video Save or Video Assemble), video_path, and info - JSON carrying the request id, the cost estimate, saved files and fal's full answer. Video only, no audio track.

Install

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

ComfyUI Manager users can skip that with Install via Git URL. Restart, then:

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

in the .env next to run_nvidia_gpu.bat. Menu path: arkennemasis/fal/Video/MiniMax.

Traps and habits

The .env location is the recurring one - beside the launcher, or in the ComfyUI folder. Nowhere else, and no key box on the node.

The habit worth building: prompt iteration on hosted video is where a $20 afternoon disappears. Run the cheap rung first, keep the wording that works, then re-render at 1080p. And if you're stringing several fal nodes together, leave max_concurrent at 1 unless you have a reason - the default sequencing is what keeps a canvas full of paid nodes from slamming fal's rate limits and failing half of them.

Last thing: reuse_identical_run will hand back the previous clip when nothing changed. That's a feature. Flip it off when you genuinely want a new sample from the same prompt - the same text does not always give the same video, and the tooltip says so plainly.