arkennemasis fal · MiniMax Hailuo 02 Standard Image to Video · $0.017-0.045/s

This is the middle of the Hailuo ladder, and it's the one most people should actually use. Feed it a picture, pick a resolution and a duration, and you get a clip: 768p at roughly $0.045 a second (a 6-second shot is about 27 cents) or 512p at roughly $0.017 a second (about 10 cents). Fast tier caps at 512p; Pro tier is 1080p for a flat 48 cents. Standard gives you both ends and everything in between.

It's also the node where the two useful Hailuo extras live side by side: a duration choice and an optional end frame. And since Hailuo is a closed hosted model with no weights anywhere, the endpoint is the only way in - the local-vs-API argument doesn't apply, only the price-per-quality one.

Inputs

• image - the first frame. Standard-tier image-to-video is compositionally honest: the output tracks your picture, which is why a clean, unambiguous start frame matters more than the prompt.
• prompt - required. Motion and camera direction.
• resolution - 768P (default) or 512P. This is the price dial, roughly 4.5 cents per second against 1.7.
• duration - 6 or 10 seconds, default 6. Note that 10-second clips aren't supported at 1080p; on this endpoint both options are fine, but the constraint is worth remembering when you jump to Pro.
• end_image - optional, and the interesting one: supply a last frame and the model has to interpolate toward it. Two-frame control is the cheapest reliability upgrade available in AI video.
• prompt_optimizer - default true. The model rewrites your prompt. Turn it off when you need your exact words.
• Plus max_cost_usd, reuse_identical_run, extra_json, max_concurrent.

What it costs

Billed per second of generated video, at the rate your resolution picks. The pack builds both its live price badge and its pre-run estimate from the same rule, and the estimate uses the duration box, so a 10-second 768p shot shows up as about 45 cents before you commit. max_cost_usd (default $20) blocks the run before the upload if the estimate exceeds it; at these numbers it only fires when something upstream is wrong, which is a good guard.

reuse_identical_run (default on) matters more in video than in stills - repeating a graph while tuning a downstream node gives you back the clip already paid for, rather than re-buying it. Turn it off deliberately for a fresh generation from the same settings. max_concurrent (default 1) keeps paid fal calls sequential across the whole run; that's the setting that keeps a 20-clip batch under fal's rate limits.

Outputs

video - ComfyUI's VIDEO type, into a save node or downstream tooling - video_path as a string, and info, the JSON with the request id, the cost estimate, the saved files and fal's full answer. Files land in output/fal/fal-ai_minimax_hailuo-02_standard_image-to-video/.

Install

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

Or install comfyui-arkennemasis from ComfyUI Manager. Restart, then:

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

in the .env file next to run_nvidia_gpu.bat. No models, no dependencies, no key widget on the node - the key lives in that one file, which is why it can't end up inside a workflow you hand to somebody.

Practical wisdom

Test at 512p, deliver at 768p. The motion is largely the same at both, and the price difference is nearly threefold, so the efficient loop is: 512p, six seconds, iterate the prompt and the start frame, then re-run the winner at 768p. And note that changing the resolution is changing the inputs, so reuse_identical_run won't quietly hand you the cheaper clip back when you upgrade it.

If clips look mushy, check that your start frame isn't smaller than the output you asked for - and that it isn't a busy composition the model has to reinterpret. If results feel like they drifted from your text, try prompt_optimizer off. If a job fails, the fal nodes still queued in that run don't start, Cancel cancels on fal, and the free Recover Result node recovers a finished job by id if ComfyUI restarts mid-wait.