arkennemasis fal · Flux 3 Text to Video · $0.17-0.29/s

Some context first, because it explains why this node is a paid call and not a download. FLUX 3 is Black Forest Labs' multimodal turn - one model over image, video, audio and robot action - and it shipped in July 2026 as availability tiers, not checkpoints. Video is early access: API plus private weights. The open-weight Dev backbone is a promise with no date attached. Meanwhile the community reaction to the showcase was the same one BFL got for Flux 2, and flux-2.md doesn't soften it: lot of hype, "you're counting on the hype from our open-source community", shown as the paid API.

So: this is the paid API. It's also genuinely good, has native synchronized audio, and runs up to 20 seconds in one generation. ArkFal_blackforestlabs_flux_3_text_to_video is the plainest entry point - prompt in, video out, $0.17/s at 720p or $0.29/s at 1080p.

What you set

• prompt - the shot. Say what happens, in order, with a camera instruction if you care about the camera.
• resolution - 720p (default) or 1080p. That's the whole pricing decision, right there: 1080p costs ~70% more per second.
• duration - auto lets the model choose, or pin it. Pin it when you have a timeline to hit; auto when you don't.
• generate_audio - on by default. Dialogue, SFX, ambience come back with the clip. Turn it off if you're going to lay your own track over it, though it doesn't change the price.
• aspect_ratio - auto by default, which is honest about the fact that a text prompt doesn't really imply a shape.
• safety_tolerance - 0 strictest, 4 most permissive, default 2.

The optional controls are the pack's shared fal set: max_cost_usd (default $20 - and this is the node that most deserves a tighter one; a 20-second 1080p clip is $5.80, so three of them clear the cap and stop the run before anything is uploaded), reuse_identical_run, extra_json, max_concurrent.

Outputs: video, video_path, the seed, and info.

Cost, honestly

Seconds are the unit and the duration box is what the estimate reads. When duration is auto, the badge can't know what the model will pick, so it shows a rate in the browser and a real dollar figure at run time. Here's the arithmetic worth internalising: 10 seconds at 1080p is $2.90. A six-shot scene at that quality is nearly $18. That's not a reason to avoid it - it's the actual price of a closed frontier model, and the community's complaint that these are expensive is really a complaint that the free-lunch era is over - but it is a reason to draft first.

The draft version of this exact node costs $0.06/s, and the draft feeds draft_enhance at $0.29/s. Iterating in drafts and finishing once is the difference between a $4 experiment and a $20 one.

Install

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

Restart after that, or install through ComfyUI Manager by searching arkennemasis. Then:

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

in the .env that sits next to run_nvidia_gpu.bat (the ComfyUI folder on a normal install). The nodes read the key from there and nowhere else - there's no key widget, which is on purpose: a typed key would be stored in every workflow you save or share.

Where people get burned

The .env in the wrong folder is the classic. If the node reports no key, that's almost always it, or the file being saved as .env.txt.

Second: this is a cloud call. Your prompt leaves the machine, and the model's filter is the model's filter - no local LoRA, no abliteration, no arguing. That's the trade external-api-nodes.md documents for this whole category, and it doesn't change because the wrapper is neat.

Third, the practical stuff: results land in output/fal/blackforestlabs_flux-3_text-to-video/, every request is logged to output/fal/_requests.jsonl, rate limits and 5xx drops are retried automatically with backoff, and ComfyUI's Cancel cancels the job on fal too. If your ComfyUI restarts while a clip is rendering, the work isn't lost - fal Recover Result picks it up by request id, free.