arkennemasis fal · Flux 2 Flash · $0.005/MP

Half a cent per megapixel. A 1024×1024 render out of Flux 2 through this node costs about $0.005 - under a penny - and that makes it the cheapest sensible way to iterate on a Flux 2 prompt before you commit the good model to it.

This is the speed tier of the family, not a different model. Same realism push, same text rendering, same "it writes legible signage" trick. What's missing compared to the full Flux 2 node is the knobs: no num_inference_steps, no acceleration dropdown. The endpoint handles sampling itself, and you get the prompt, the size and a few switches. That's the trade. You're buying cheap and fast, and you're giving up the ability to trade wall time for quality inside the node.

The inputs

prompt - the model's encoder reads like an LLM chat turn, so write sentences. "A studio product shot of a matte black water bottle on a pale grey seamless backdrop, softbox from the left" beats a pile of comma-separated tags, and ((emphasis:1.4)) weighting does nothing at all.

image_size - presets (landscape_4_3 is the default) or custom, which switches on image_size_width / image_size_height. Width and height are the actual cost driver here: 1024×1024 is 1MP, 1920×1080 is roughly 2MP, and the bill follows.

num_images - this is the one to use properly. At half a cent a shot, asking for four at once is the whole point of a draft stage, and the node returns them as one batch on the images output.

guidance_scale (2.5) and seed (-1) - leave them unless you're reproducing something.

enable_prompt_expansion - off. It hands your prompt to an LLM for rewriting. Great if you typed six words and want twenty; terrible if you already know what you're asking for.

enable_safety_checker - on, and it flags rather than errors: a refused image comes back as a black square. Turning it off requires authorization on your fal account, and unauthorized requests are checked regardless.

output_format - png, jpeg or webp. PNG for anything you'll edit further.

Outputs

images (IMAGE, batched if you asked for several), image_paths (STRING - where they landed), seed (INT), prompt (STRING, what the model received after any expansion), timings, has_nsfw_concepts, and info - JSON with request id, cost estimate, saved files and fal's full answer. Wire images into SaveImage, a contact-sheet node, or straight into the Flash Edit sibling if you're chaining an edit pass.

Install

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

Restart ComfyUI; the node lives under arkennemasis → fal → Image → Flux. ComfyUI-Manager can grab the pack by Git URL. The pack's only pinned dependency is httpx, which belongs to its ChatGPT/Codex nodes - the fal provider uses Python's stdlib, so this node installs clean on any working ComfyUI.

Then the key, one line in the .env by run_nvidia_gpu.bat:

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

Nothing on the node asks for it, deliberately - a key that lives in a file can't end up baked into a workflow you paste on Discord.

Where people get burned

Expecting it to be the good model. Flash is the draft tier. If your hands keep coming out wrong, that's not a Flash bug - it's why the full Flux 2 node exists. Draft here, re-run the keeper there.

Forgetting the megapixel math. Price scales with area, so a 2MP render is twice a 1MP one and num_images multiplies it again. Four 2MP drafts is about four cents, which is still cheap - just don't build a 200-image batch by accident.

Wondering why the second identical run was instant and free. reuse_identical_run is on by default: same settings, same inputs, files still in output/fal, so the node hands them back rather than paying twice. Turn it off for a new take.

Tripping the cap mid-batch. max_cost_usd defaults to $20 and is checked before upload, so an over-budget run stops for free. Lower it to a dollar or two while experimenting and you'll never have a surprise. If a fal request does fail, the node retries rate limits and gateway errors itself; jobs queued behind a failure aren't started, so you don't get billed for a cascade.