arkennemasis fal · MiniMax H3 Max Styles Hand Drawn · $0.08/s

Pencil-look motion without the style lottery

The hand-drawn style is baked into this endpoint, so you write what happens and the model draws it in that look. No style keywords, no drift between shots, no discovering that your carefully-tested prompt produces oil painting on a Friday.

Where it earns its keep is animatics. You have a storyboard, a voice track, and a client who needs to see the timing before anyone animates anything. Local tools can do this, but holding a consistent hand-drawn look across eight shots while also wrangling an open video model is a two-weekend project. This is a per-second rental by comparison, and the output already includes audio.

$0.08 per second at 768p, audio included - five seconds is $0.40, the maximum fifteen is $1.20.

What's in and what isn't

Text prompt in, 768p video with sound out, 5–15 seconds. Optionally a first-frame image.

Inputs that matter

• prompt - describe the scene, action, camera and sound. The style is applied for you; the words are for content. If you want pencil texture and a specific line weight, you're in the wrong node - this is a style endpoint, not a style editor.
• image - optional first frame, and the tooltip's advice is the useful part: use pencil artwork. A hand-drawn or storyboard frame anchors the whole clip in the same medium; a 3D render as first frame makes the model convert, and conversion is visible.
• duration - seconds, default 5.
• aspect_ratio - 21:9, 16:9 (default), 4:3, 1:1, 3:4, 9:16. Ignored if you supply an image, since the image's shape wins.
• seed - -1 by default; set it to repeat a good take.

Plus the guards every fal node in the pack carries: max_cost_usd (default $20, checked before anything is uploaded - the run stops, it doesn't bill), reuse_identical_run (the same inputs return the already-paid file rather than a new charge), extra_json and max_concurrent (default 1 paid call at a time across the graph).

Outputs

video → SaveVideo, PreviewVideo, or Video Assemble if you're cutting shots into a sequence. video_path, seed (the resolved seed - the only way to reproduce a clip you left to chance) and info (request id, estimate, saved files, fal's raw answer).

Install

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

Restart ComfyUI; the node appears under arkennemasis/fal/Video/MiniMax. Key in the .env beside run_nvidia_gpu.bat:

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

Where people get burned

Feeding it a clean rendered frame. The single most common cause of "it doesn't look hand-drawn" is a first frame that isn't hand-drawn. Give it a pencil test frame, or omit the image entirely and let the endpoint decide.

Two subjects in one prompt. Sketch styles simplify; when a prompt asks for a crowd doing three things, the simplification turns to mush. One action per clip reads best.

Assuming audio is extra. It isn't - it's part of the per-second price, so describing sound in the prompt costs you nothing to try. Dialogue in a pencil-look clip is a legitimate use.

Comparing to Sparkle or SVT-style local pipelines. That's an entirely different amount of control and effort. This node is the fast path to a look; the local pipelines are the slow path to your look.

Forgetting what's on the other end of the wire. Your prompt and any first frame go to fal. For a client project under NDA, that's a decision to make on purpose.