Flux Generate 1.1 [Pro, Ultra, Raw]

The name makes it sound like a big local model. It isn't. FluxGenerate is a remote control for Black Forest Labs' hosted Flux API - the "Pro" tier of Flux that BFL has never released as weights and sells by the image. You type a prompt, the node uploads it to api.us1.bfl.ai, polls until the job finishes, and drops a finished image back into your workflow as a normal ComfyUI tensor. Your GPU just sits there watching. It doesn't even need one.

That's the whole appeal. Flux Pro is the un-distilled base model - the one Dev is guidance-distilled from - and it only exists behind BFL's paywall. If you want Pro-tier quality for a hero image, a commercial license you can actually rely on, or just a break from babysitting a 20GB checkpoint download, this node is the sanctioned way to reach it. The catch, obviously: it's paid, it needs an API key, and it only works when the BFL servers say so.

How it works

Under the hood the node is embarrassingly simple - a few hundred lines of requests. It POSTs your prompt to /v1/flux-pro-1.1 (in pro mode) or /v1/flux-pro-1.1-ultra (in ultra / ultra_raw mode), then polls /v1/get_result with exponential backoff until the image is ready. That polling loop is why a generation visibly blocks your queue for 30 seconds to a couple of minutes; there's nothing asynchronous about it. No model files, no VRAM, no sampling settings on your side - all of that happens on BFL's hardware.

The inputs that matter

• model - pro, ultra, or ultra_raw. Pro is the standard 1.1 endpoint. Ultra is BFL's higher-quality variant, and ultra_raw flips on the "raw" toggle, which skips some of BFL's default polish for a less processed, more natural look. This is the one input that changes the request endpoint, so pick deliberately.
• prompt - plain English, multiline. Flux is famously prompt-adherent, and that carries to the API.
• aspect_ratio - a dropdown of 7 presets (21:9, 16:9, 4:3, 1:1, 3:4, 9:16, 9:21). Note the mechanism: in pro mode the node converts your pick to fixed pixel dimensions (16:9 → 1408×800, 21:9 → 1408×608). No free-form resolution.
• safety_tolerance - 0 (strictest) to 6 (no moderation), default 6. Don't over-read this: moderation and censorship are enforced on BFL's servers regardless; the dial just decides how much rope they give you.
• seed - set it for reproducibility, leave at -1 for a random result.
• prompt_upsampling - when on, BFL auto-expands your prompt into something more creative. Can be magic for vague ideas, mildly annoying if you wrote something specific.

There are two optional image inputs: image_prompt (a ComfyUI IMAGE to remix) and image_prompt_strength (0–1 blend, default 0.1). One quirk worth knowing: image_prompt_strength is only sent to the API in ultra/ultra_raw mode - in pro mode the node sends the image but ignores the strength value.

Outputs

You get two: IMAGE (wire it straight into a SaveImage) and a STRING that's just a random per-session version id the author stamps on every output. It's not metadata, it's not a URL, it's noise. Ignore it.

Install

The pack ships no requirements.txt - it only needs requests and Pillow, both already bundled with ComfyUI, plus no model downloads of any kind. So install is just the clone:

cd ComfyUI/custom_nodes
git clone https://github.com/Altair200333/ComfyUI_Flux_1.1_PRO

Then the step everyone forgets, because the README buries it and config.ini is gitignored so a fresh clone won't have one: create it in the extension folder.

[API]
API_KEY=your_api_key_here
BASE_URL=https://api.us1.bfl.ai

Restart ComfyUI and you'll find the three nodes under the BFL category. (Or skip the clone and search "Flux Pro Nodes for ComfyUI" in ComfyUI Manager.)

Gotchas

If you queue and get a FileNotFoundError about config.ini, that's the missing key file - the node refuses to run without it. If a job just sits there, that's usually the polling doing its job; give it a couple of minutes before assuming it's hung. And if jobs start erroring out mid-session, check your BFL account credits first, not your workflow. The obvious trap for beginners: this is a paid API, so every queue hit is a few cents out of your BFL balance. Don't use it for seed-slider experiments - that's what local models are for. For the polished, license-clean hero shot, though, it's hard to argue with the results.