Flux Dev (fal)

If you want FLUX.1 Dev output but your GPU chokes on a 12B model, this is the shortcut. FluxDev_fal doesn't run Flux on your machine - it ships the prompt to fal.ai, renders it on their hardware, and hands you back a finished IMAGE. No 12GB-of-VRAM minimum, no downloading the transformer, T5, and VAE, no fp8 juggling. You pay fal a few cents per image instead of paying for a bigger card. That's the whole trade.

Flux Dev is the model that dethroned SDXL as the open-weights quality bar in 2024 - the tier where hands, anatomy, and text-in-image finally worked, thanks to its T5 text encoder parsing full sentences instead of tag soup. Locally it's a heavy, non-commercial-licensed beast. Wrapped as an API node it's just a box you type a sentence into.

How it works

Under the hood the node calls fal's flux/dev endpoint. Your image never touches a diffusion model on your box; the node uploads your parameters, fal queues the job, and the result comes back as a native ComfyUI image you can wire straight into a Preview or Save Image node. Because it's an API, the pack keeps a persistent result cache - ask for the exact same prompt and settings twice and the second one is free and instant. Flip force_rerun (available on the generated nodes) when you actually want a fresh roll.

The inputs that matter

Most of the schema is the usual Flux ladder, and the defaults are sane. The two you'll actually reach for:

• prompt - write in plain English sentences. Flux's T5 encoder rewards natural language; a comma-salad of tags underuses it.
• guidance_scale - defaults to 3.5, which is also the setting the community blames for Flux's "same face" look. Drop it to 1.8–2.0 for people and portraits; push it to 4+ only when you're stacking a realism LoRA.

Beyond those: num_inference_steps (28 is the standard; more rarely helps), image_size presets plus width/height for custom dims, num_images to batch up to 10, seed (-1 = random), and enable_safety_checker. Output is a single IMAGE socket.

Installing it

Grab it through ComfyUI Manager - search ComfyUI-fal-API, install, restart. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
pip install -r ComfyUI-fal-API/requirements.txt

Then restart ComfyUI. There are no model downloads - that's the point of an API pack - so the dependencies are light (mainly fal-client). The one thing you must do is set your key: grab it from fal.ai/dashboard/keys, then either copy config.ini.example to config.ini inside the node folder and paste it there, or export FAL_KEY=your_key. The node shows up under the FAL category.

Where people get burned

The first surprise is that it costs money - every render is a real fal API charge, not a free local generation. Budget for it, and lean on the result cache so you're not re-paying for identical calls.

The second: it's licensed for what you generate, but not for hosting the model - that distinction lives with BFL's weights, and here fal is doing the hosting, so you're free to sell the images you make. Don't expect NSFW, though; Flux Dev's endpoints run BFL's content policy, and a blocked request comes back as a visible fal error rather than a silent blank. If a call fails, read that error text - it usually names the exact parameter or policy that tripped.

If you're happy trading a little quality for speed and cost, its sibling FluxSchnell_fal runs the distilled 4-step tier for less per image. Use Dev for the final, Schnell for the fishing.