Fal API Flux Dev

Fal API Flux Dev doesn't run Flux. It sends your prompt over the wire to fal.ai, lets their GPUs run the actual FLUX.1 [dev] checkpoint, and hands the finished image back into your ComfyUI graph a few seconds later. If you came here expecting a model loader, this is the opposite: no checkpoint file, no VRAM budget, no local queue. You're calling someone else's server.

That's worth sitting with for a second, because it changes how you should think about the node. FLUX.1 Dev is Black Forest Labs' guidance-distilled workhorse - the open-weights tier sitting between the throwaway-fast Schnell and the API-only Pro - and normally its license is the annoying part: non-commercial for the weights themselves, with a real fee if you want to self-host it commercially. None of that is your problem here, because you never download the weights. fal.ai holds the commercial arrangement and rents you API calls; you're a customer, not a host. What you are paying for, one generation at a time, is fal's metered compute - this isn't free once your GPU is warm, it's billed every single run.

Mechanically it's about as simple as these wrapper nodes get: fill in a prompt and a few generation knobs, the node makes the API call, and out comes an IMAGE tensor you wire straight into Preview Image or Save Image, same as any local sampler's output.

What you actually set. prompt is the only required text field, and it's worth writing it like a sentence rather than a tag list - Dev's T5 text encoder was built for natural language, and stacking comma-separated tags the SD1.5 way genuinely underuses it. width and height default to 1024×1024 in steps of 8, which is Flux's native comfort zone; it'll happily go up toward 1536×1536 or beyond if you want more detail per generation. num_inference_steps defaults to 28 (1–100) - Flux doesn't need SDXL-style step counts, and past the high 20s to low 30s you're mostly burning fal credits for diminishing returns. guidance_scale defaults to 3.5, and this is the one dial worth second-guessing: it's the same value widely blamed in the community for Flux's homogenized "same face" look on people. If portraits keep coming out generic, try dropping it to 1.8–2.0 before you touch anything else. num_images lets you batch 1–4 per call (each one billed), and enable_safety_checker toggles fal's content filter, on by default. seed is optional - leave it at 0 for a fresh result each run, or pin whatever value you got back once you land on something worth iterating against.

The only output is IMAGE.

Installing it. Easiest path is ComfyUI Manager: search "ComfyUI-Fal-API-Flux," install, restart. Doing it by hand means cloning into custom_nodes - worth flagging that the README's own manual-install command has a copy-paste bug, pointing at your-username/ComfyUI-Fal-API-Flux instead of the real repo; use https://github.com/yhayano-ponotech/ComfyUI-Fal-API-Flux.git, or just let Manager sidestep the whole issue. Either way you then pip install -r requirements.txt if you went manual, and - this is the step people miss - create a config.ini in the pack's own folder with a [falai] section and your api_key. This isn't wired into any of ComfyUI's usual credential UI; it's a plain text file the node reads on load, so if the node silently fails or throws an auth error, that file is the first thing to check, followed by whether your fal.ai account actually has FLUX.1 access enabled.

Where it actually breaks. Almost every real issue traces back to the API boundary rather than anything ComfyUI-specific: a missing or malformed config.ini, a fal account without Flux enabled, or a network hiccup that shows up as a vague error in the ComfyUI console rather than in the node itself - the README's own troubleshooting advice is to read that console output carefully, because that's where fal's actual error message lands.