FLUX 1.0 [dev]

Here's the awkward one: FLUX 1.0 [dev] is the API version of a model you can already run locally for free. Dev is the open-weights workhorse that took over from SDXL - 12B parameters, T5-XXL text encoder, guidance-distilled, non-commercial license. This node wraps the exact same model on BFL's servers and charges you per image for the privilege. So before anything else: is that worth it?

Sometimes, genuinely yes. Dev's full fp16 stack is bigger than most consumer VRAM, and even quantized it wants serious hardware. If you're on a laptop, a cloud box you don't control, or just don't want the 25GB of downloads, this node gives you Dev's quality with zero VRAM pressure - your GPU idles, BFL's servers do the work. It's also the most controllable node in the pack, which makes it a decent API on-ramp. But if you own a big enough card, the local workflow is free and this node is a convenience tax.

How it works

Same mechanism as the whole pack: the node sends your prompt (and optional reference image, base64-encoded) to the v1/flux-dev endpoint, polls until ready, and returns a standard ComfyUI IMAGE. Notable extra: prompt_upsampling defaults to true, meaning the API rewrites and enriches your prompt server-side before generating. That's why a short sentence here often comes back with better adherence than the same sentence on local dev.

Inputs that matter

• prompt (required) - natural language, not tag soup.
• guidance (default 2.5, range 1.5–5) - matches the API's guidance scale. Community lore on local dev says the old 3.5 default is too hot for faces; 2.5 is already the saner end.
• width / height (up to 1440, step 32) - explicit resolution control, unlike the ultra node.
• steps (10–100, default 50) - Dev is guidance-distilled, so it converges well before 100; you're mostly paying for overkill past ~40.
• interval (1–10, default 1) - the API's guidance interval. Leave it alone unless you know why you're changing it.
• image_prompt (optional IMAGE) - for image-remix generations.
• prompt_upsampling (default true) - flip off if you want the API to leave your wording alone.
• safety_tolerance (0–6, default 2).

Output is one IMAGE tensor - Save Image, or feed it into whatever comes next.

Install

Part of the official Black Forest Labs API Nodes pack: ComfyUI Manager → search the pack title, or

cd ComfyUI/custom_nodes
git clone https://github.com/black-forest-labs/bfl-comfy-nodes

restart, then set your key (export BFL_API_KEY=..., a bfl_api_key.txt in the pack folder, or the api_key_override field). Key from api.bfl.ml. No model files, no extra pip packages - the pack runs on ComfyUI's stock dependencies.

Where the trap is

The trap isn't technical, it's financial: this node bills per image for a model you might legally run free. Test with two or three cheap images before building a whole workflow on it. Also remember Dev's license is non-commercial, and the API path doesn't change that - using this node to dodge the license doesn't work, and the moderation is tighter than your local setup. If you just want the best look with no dials, that's the ultra node. This one is for people who want Dev's quality on hardware that can't carry it.