Flux 2 Klein 4B (BFL)

Flux 2 Klein 4B (BFL) is the little Apache-2.0 sibling in the Flux 2 family, and through this node it's available without a single local download. Klein is BFL's size-distilled consumer variant of Flux 2 - the 4B model is the one that's fully open-weights (Apache 2.0, the only commercially-usable Flux 2 checkpoint), famously small enough to run in roughly 13GB of VRAM locally. The irony: this node lets you run it without even the 13GB, because the compute happens on BFL's servers.

So who's this node actually for? Two groups. First, the person with a weak GPU who wants Klein 4B's quality without the local install - the same pitch as every node in the pack. Second, and more interestingly, the person who wants the commercial model but wants BFL to handle it: on the API you get the same model as the open weights, with the licensing question moot because you're renting inference, not redistributing weights.

How it works

Standard pack flow: prompt, optionally up to four base64 reference images (Klein is the smaller model - it tops out at four, where the bigger tiers take eight), POST to flux-2-klein-4b, poll get_result?id=..., return an IMAGE. It's the Klein treatment of the same multi-reference generation-and-editing concept Flux 2 is known for, scaled to the small model.

The inputs that matter

• prompt - the description.
• input_image … input_image_4 - up to four references, base64 from Image to Base64 (BFL). This is the character/object consistency move, Klein edition.
• width / height - optional (≥64), default lets BFL decide.
• safety_tolerance - 0–5, default 2. Moderation; Flux 2's filtering is in the weights, this just sets tolerance.
• output_format - jpeg (default) or png.
• Optional seed, webhook_url/secret, config.

Installing it

One pack, one install, no model files:

cd ComfyUI/custom_nodes
git clone https://github.com/gelasdev/ComfyUI-FLUX-BFL-API.git

Restart, key in config.ini or via Flux Config (BFL). Only dependency is torch, already in ComfyUI.

Where people get burned

• Paying for what you could run free. If you have the ~13GB of VRAM, Klein 4B locally is Apache 2.0 and uncensored (relative to the API) - the API version is metered and moderated. This node is the right call for weak hardware or for commercial licensing convenience, not a replacement for local Klein if you can run it.
• Expecting the 9B's quality. Klein 4B is the budget tier of the family. Fine for concept work, softer on hard details than the 9B. Don't blame the node.
• Black output. The pack's universal failure tell - moderation, API error, or poll timeout. [BFL] logs print a curl equivalent per call; replay it to see the raw API response.

The honest framing: Klein 4B via API is a convenience product, not a capability unlock - the model is genuinely open-weights and genuinely small. Where it shines is the 4GB-GPU laptop owner or the company that wants a commercially-usable Flux 2 tier without hosting anything.