arkennemasis fal · MiniMax H3 Max Turbo Text to Video · $0.025-0.08/s

Every text-to-video model in the world is priced by the second now, and this node's job is to be the cheap, fast lane for MiniMax H3 - fal's post-trained turbo variant, tuned for stronger prompt adherence and better aesthetics, running on co-optimised inference. Nothing to load, nothing to download. Type a prompt, pick a resolution, pay for the seconds you asked for.

Worth being clear about why you'd use the hosted one at all: H3's open weights are real, but the community licence excludes the US, EU, UK and South Korea from running them, and the model is a 33B monster with no published consumer hardware floor. This endpoint is globally available and has no VRAM requirement. That's the whole case for it.

How a run goes

The node reads FAL_KEY from your install's .env, checks the inputs, estimates the cost and refuses if it's above max_cost_usd, uploads nothing much (you're sending text), submits - that's the one billed call - waits with live status, then drops the file into output/fal/minimax_h3-max-turbo_text-to-video/ and shows it on the node. Failures before the submit step never bill. ComfyUI's Cancel cancels the job on fal too, so you're not leaving a render running out of sight.

Inputs worth thinking about

prompt and duration (5–15s) are the two you'll touch every time. Longer is billed linearly, so five seconds is the sane place to iterate and fifteen is for the take you've already validated.

resolution - 480P, 768P or 1080P, default 768P. Read the tooltip before you reach for 1080P: it's latent refinement from a native 768P source, not a native 1080p generation, and it costs the most. If you're delivering at 1080p anyway, it can still be the cheaper road than generating at 768p and running a separate upscaler, because you're not paying another API.

aspect_ratio defaults to 16:9, which is why this node has the widget while the image-to-video sibling doesn't - there's no input picture to infer a canvas from.

prompt_expansion_mode is the quiet one. balanced spends about a second rewriting your prompt before generation; quality can spend up to ~30 seconds on a richer rewrite; disabled sends exactly what you typed. The node hands back the expanded_prompt output so you can see what was actually used - if a video went somewhere weird, check that string before you blame the seed.

target_audio pins an audio clip (2s minimum, 15MB max) over the generated soundtrack. H3 always produces audio, so this is replacement, not addition - useful when you already have narration, and the thing to know if you were hoping for a silent clip. You don't get silence from this model.

The rest is pack-wide plumbing: max_cost_usd ($20 default cap, 0 off), reuse_identical_run, extra_json for fal fields the node has no box for, and max_concurrent (1 by default, meaning one paid fal call at a time across the entire run).

Outputs are video (VIDEO - wire it to Save Video or the pack's assembly nodes), video_path, timings, expanded_prompt and info, the JSON with request id, estimate, saved files and fal's raw answer.

Install

Manager → search arkennemasis, or clone it and install the one dependency:

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

Then, in the .env next to run_nvidia_gpu.bat:

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

Restart. No key field on the node, by design.

Where people trip

• Same output twice, different seed. reuse_identical_run is on by default and returns the files you already paid for when the settings and inputs match. Flip it off for a fresh take; leave it on for cheap re-runs after a graph tweak downstream.
• The 1080P surprise. It's the priciest option and it's an upscale path. Check the tooltip, then check max_cost_usd.
• A wall of fal nodes that trickles. max_concurrent: 1 on every node means they queue. Raise it where you want parallel calls - and note that once any fal node fails, the queued ones are never started, which is a feature when you're spending money.
• Restart mid-render. Grab the request id from output/fal/_requests.jsonl and pull the finished job down with fal Recover Result - free, because nothing is submitted.