arkennemasis fal · Flux Dev · $0.025/MP

FLUX.1 [dev] is the 12B model that displaced SDXL within months of its August 2024 release, and it's still the most predictable thing in open image generation. The launch-thread reaction was "What the fuck. It's insanely good", the prompt adherence has held up across two model generations, and its known weaknesses are equally well documented: a slightly plastic base look, and hands that multiply when a full body fits in frame.

You can run it locally. That's the thing about this node - it isn't the only way to reach Flux Dev, or the cheapest over a week of heavy use. It's what you reach for when you want 200 images rendered while you do something else, or when the machine you're on has no usable GPU at all. $0.025 per megapixel: about two and a half cents an image at 1024×1024.

Is hosted worth it here?

Compare honestly. Locally, Flux Dev is free after electricity and you own the loop. On fal, you're paying per image and your prompt goes to someone else's server - that's the shape of every API node in ComfyUI, and it's the reason a lot of people who could run this locally still wouldn't.

Where the hosted version wins: batching a client's list overnight from a laptop, handing a workflow to a colleague without a 4090, and the fact that the endpoint's settings are the model author's own rather than the ones you guessed at 1am. Where it loses: iteration. Twenty quick test renders on your own card cost nothing; the same twenty here cost fifty cents, which is nothing too, but the round trip is the tax.

The settings

prompt - plain sentences. Flux's T5-derived encoder wants natural language, not tag soup.

guidance_scale defaults to 3.5, and it's Flux guidance, not CFG. The distinction bites people constantly: applying SDXL's CFG logic here doubles generation time by running the model twice, which isn't what this parameter does. Higher guidance (5–8) buys prompt adherence at the cost of that plastic sheen; 3.5 is the community's settled number.

num_inference_steps defaults to 28 - the launch-era convention landed at 20–30, so leave it. seed at -1 for a random draw; pin it when you're comparing prompt versions against a fixed composition. image_size is a combo (landscape_4_3 default, custom with image_size_width / image_size_height), num_images goes up to 4 in one call, output_format defaults to jpeg here.

acceleration defaults to none and is the field people miss. It's a speed setting - the higher you push it, the faster the generation. It's the difference between a four-second render and a considerably quicker one, worth trying when you're exploring and worth leaving alone when you're delivering.

enable_safety_checker stays on; disabling it needs authorization from fal at the account level, and unauthorised requests get checked regardless.

Outputs

images and image_paths (files in output/fal/fal-ai_flux_dev/) for the pipeline, then the diagnostics: timings, seed, prompt, has_nsfw_concepts, and info - request id, cost estimate, saved files and fal's complete answer as JSON.

Keep seed and prompt wired somewhere. When you land a good one, those two strings are the only way back to it.

Install

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

Restart ComfyUI, or use ComfyUI-Manager → Install via Git URL. The pack declares exactly one dependency, httpx, and it exists for the Codex nodes; the fal provider needs nothing beyond Python's standard library. No checkpoints, no VAE, no 12GB download - the model is on fal's side.

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

goes in the .env next to run_nvidia_gpu.bat (portable build) or in your ComfyUI folder. There is no key field on the node. That's deliberate: a key in a widget gets written into every workflow JSON you share.

The gotchas

reuse_identical_run is on by default, and on this node it's a feature you'll notice immediately. Re-queue an unchanged graph and you get the cached file back with nothing billed - great when you're re-running a workflow to test the rest of the graph, startling when you were clicking Queue for a fresh random image. Turn it off for that.

Estimate your batch: four images at 1536×1536 is roughly nine megapixels, about 22¢ per queue. max_cost_usd (default $20) checks that estimate before anything is uploaded, and the node's price badge shows the number as you change settings.

One last honest note on licensing. FLUX.1 [dev] ships under BFL's non-commercial licence, and the v2.0 revision covers "FLUX.x [dev]" including Fill, Canny, Depth, Redux and Kontext. If you're generating for a client, that's your problem to settle before you deliver, not fal's - read the current licence rather than relying on a forum summary.