arkennemasis fal · Flux Schnell · $0.003/MP

Flux Schnell is Apache 2.0 and runs in one to four steps, so you could just run it locally for free. Given that, why is there a node for the hosted version at $0.003 per megapixel?

Three reasons, and they're all legitimate. You don't have a GPU. You want it in a workflow that runs somewhere else - a server, a headless machine, a batch job. Or you're throwing away drafts by the hundred and would rather not wait for your card to chew through them. At a third of a cent per megapixel, a 1024×1024 image costs three-tenths of a penny. You can generate a hundred candidate compositions for thirty cents and pick one to redo properly.

Being honest about the flip side: Schnell is the distilled tier, and the two things people wanted from Flux - variety in faces and trainability - are exactly what the distillation cost it. It's too step-distilled to fine-tune directly, which is why BFL could afford to release it under Apache 2.0 in the first place. Use it for layout, composition, thumbnails, and "does this prompt even work" sanity checks.

Inputs

• prompt - natural language. Flux's T5 encoder reads sentences; write them.
• num_inference_steps - default 4, max 12. Schnell's whole trick is running at 1–4 steps, and the pack exposes the dial rather than pretending it doesn't exist. Going up to 12 mostly costs time.
• image_size - a preset list (landscape_4_3 default, plus square_hd, square, the portrait presets, and custom), with image_size_width/image_size_height for the custom case. Billing is per megapixel, so this is the price dial: square_hd is 1 MP, a 1536×1536 is over 2 MP and rounds up to three cents.
• guidance_scale - default 3.5, range 1–20. Yes, Schnell has a guidance box here; the local folklore that Schnell has "no guidance parameter" is about the open-weights inference defaults, not this endpoint.
• acceleration - none / regular / high. Speed versus quality polish; start at none.
• num_images (up to 4), seed (pin it with a real number once something works), output_format, enable_safety_checker.
• The shared money widgets: max_cost_usd, reuse_identical_run, extra_json, max_concurrent.

What comes back

images and image_paths, files written under output/fal/fal-ai_flux_schnell/, plus seed as an INT, prompt, timings, has_nsfw_concepts and info - the JSON with fal's request id, the cost estimate and the full answer. If you set acceleration or change size and want to know what it actually cost, info is where you look.

Install

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

Or install comfyui-arkennemasis from ComfyUI Manager and restart. Then put the key line in the .env that lives next to run_nvidia_gpu.bat:

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

There's no key widget on the node deliberately - a typed key would be saved into every exported workflow. And unlike the pack's ChatGPT-based nodes, nothing here needs httpx: the fal wrapper talks HTTP with the standard library.

The safety-checker quirk

Worth knowing before you're surprised by it. If enable_safety_checker is on and the model flags the image, you don't get an error - you get a black image. And turning the checker off requires account authorisation on fal's side: unauthorised requests are checked anyway, whatever the widget says. So a black square at three-tenths of a cent isn't a bug in the node; it's a moderation result with a poor user interface.

Practical notes

Two failure modes are really habits. First, reuse_identical_run is on by default, so hitting Run twice on an unchanged graph returns the same file from disk rather than paying again - which is what you want in a batch, and the opposite of what you want when you're exploring with a fixed seed. Set seed to -1, or turn the reuse switch off, if you want new pixels. Second, if you stack this node behind a LoRA or an upscaler, remember Schnell's ceiling: it's a fast composition engine, not a finisher. Generate at Schnell, then redo the winner on something with a Pro base.

Missing key, or an empty fal balance, stops the run before it bills anything - and if ComfyUI is restarted mid-job, the free Recover Result node in the same pack grabs the finished request by id instead of you paying twice.