arkennemasis fal · MiniMax H3 Max Text to Video · $0.05-0.16/s

Words in, a shot with sound out

This is the plainest node in the H3 Max family: no image to animate, no video to extend, no references to number. You write a prompt and get a video. If you're new to the family, this is the one to learn on, because the H3-specific habits you pick up here - how resolution tiers actually work, what prompt expansion does to your wording, what native audio means for your prompt - carry over to all the others.

Two things make it worth knowing about beyond "it generates video". First, it's H3 Max, fal's post-trained variant of MiniMax's open 33B omni-modal model, tuned for stronger prompt adherence and better aesthetics. Second, the audio is generated with the picture rather than added afterwards, which means describing a sound in your prompt is a real instruction, not decoration.

Inputs that matter

• prompt - the shot. Scene, action, camera, sound. H3 Max follows instructions decently, so write like a director rather than a mood board, but keep it to one shot's worth of content.
• resolution - 480P, 768P (default), 1080P. The honest reading of the tooltip: the top tier is a 1080P latent refinement from a native 768P source, so 1080P is a finishing step, not four times the detail. Pricing is quoted at $0.05/s (480p), $0.08/s (768p) and $0.16/s (1080p) as post-promotion rates - fal ran 50% launch discounts on some H3 Max endpoints through 30 September 2026, so check the node's badge rather than a blog post.
• aspect_ratio - 21:9, 16:9 (default), 4:3, 1:1, 3:4, 9:16. Note there's no auto on this node: text-to-video has no source to follow, so you pick.
• prompt_expansion_mode - disabled, fast, balanced (default), quality. The single most consequential setting on the node. disabled sends your prompt as written; fast returns in about a second; balanced decides per request; quality can spend up to ~30 seconds rewriting it into something richer. Richer is often better for a vague idea and often worse for a precise one. Learn what your prompt turns into by reading the expanded_prompt output before you decide.
• duration - seconds, default 5.
• target_audio - pin an audio clip you provide over the generated soundtrack. Needs to be at least 2 seconds and no more than 15 MB; longer clips get trimmed to the video duration (keeping the beginning), shorter ones are silence-padded, and playback speed is never changed. If you have a music bed you want used exactly as-is, this is the input.
• seed (-1 = the model picks) and enable_safety_checker (on by default).

And the guards that come with every fal node in this pack - worth learning once, because they're identical everywhere: max_cost_usd (default $20, stops the run before anything is uploaded if the estimate exceeds it; 0 removes the cap), reuse_identical_run (returns the paid result for identical inputs, even after a ComfyUI restart), extra_json (a JSON object merged into the request last, for any fal field the node has no widget for) and max_concurrent (default 1 - one paid fal call at a time across the whole graph; raise it on the nodes you want running side by side).

Outputs

video, video_path, timings (fal's stage timings - the place to look when a run felt slow), expanded_prompt (what was actually sent) and info (request ID, cost estimate, saved files, fal's full answer).

Install

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

Restart ComfyUI, then find it under arkennemasis/fal/Video/MiniMax. The key is one line in the .env next to run_nvidia_gpu.bat (portable build) or in your ComfyUI folder:

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

There's no key box on the node, deliberately - a widget value would end up embedded in every workflow you save.

Where people get burned

Thinking 1080p means sharp. It's refinement of a 768P base. If you want more real pixels, you upscale afterwards - or you accept that a paid 1080P tier and a free local upscale are doing similar work.

Letting quality expansion eat your intent. It's the default's neighbour, not the default, but people switch it on because "quality" sounds strictly better. It isn't; it's a different prompt. Read expanded_prompt whenever a result surprises you, then decide.

Drafting at the expensive tier. Five seconds at 480p is a quarter the price of five at 1080p. Motion and composition decisions belong at 480p or 768p; the final render is the only run that needs the top tier.

Forgetting that sound is generated. H3 makes its own audio, including speech-shaped sounds. If a clip is meant to have no soundtrack, that's a cleanup step (or the pack's Video Dub / Narration Length nodes, which exist because H3 cannot be asked for silence).

Restarting mid-run isn't a loss. The submit is logged in output/fal/_requests.jsonl; the job keeps running on fal and fal Recover Result collects it by request ID for free. fal · History (free) loads any completed run back into a fresh graph without paying again.