FLUX.2 [klein] 4B

FLUX.2 is Black Forest Labs' flagship, and the community's verdict on the big one was complicated: stunning on paper, but it needed 18–24GB of VRAM even quantized, and its model card read like a censorship changelog. Then BFL shipped Klein - the size-distilled answer, with the 4B rung licensed Apache 2.0 and able to fit in consumer hardware. That's what this node runs (runware:400@4): the FLUX.2 experience at a fraction of the size and, via Runware's cloud, at a fraction of the cost of running the 32B locally.

If FLUX.2 was the proof that winning benchmarks isn't winning the community, Klein 4B is the peace offering that reopened the consumer front. It's the model you point a beginner at and the one you reach for when you want FLUX.2's strong prompt-following without the hardware commitment - or, in this node's case, without the hardware at all.

How it works

imageInference with the runware:400@4 AIR. Standard mechanics - prompt and referenceImages up as base64, image back as IMAGE. What's distinctive is the sampling defaults, which reflect a distilled model's needs: steps defaults to 4 (vs 30 for SDXL) and CFGScale defaults to 3.5. Distilled models jump in few, large steps - that's the whole technique - so don't bring 30-step muscle memory. There's also an acceleration dropdown (none/low/medium/high) that trades a bit of fidelity for speed, plus a referenceImages input for FLUX 2's editing and character-consistency tricks.

The inputs that matter

• positivePrompt (required) - FLUX 2's prompt-following is its reputation; give it detail.
• width / height - 1024×1024 defaults, 128–2048, step 16.
• steps - 4 default. This is the distillation talking: more steps buys less than you think on a model trained to finish in a handful.
• CFGScale - 3.5 default. FLUX-family models run low CFG; single digits are the whole range that matters.
• referenceImages - the editing input. Wire in an image to instruct editing or enforce a subject across generations.
• negativePrompt - present, but FLUX models historically ignore it the way SDXL respects it; don't lean on it.

Installing it

Install the pack, not the model. ComfyUI Manager → search Runware → install → restart. Manual:

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

Key from runware.ai/api-keys → ComfyUI Settings → Runware API key (or RUNWARE_API_KEY). Klein 4B is hosted; nothing downloads.

Where people get burned

• 4B is not the 32B. Klein 4B is the consumer rung - great value, still a distilled small model. If you need the flagship's ceiling, the pack's other FLUX nodes point higher and charge more.
• Don't crank steps. The 4-step default is the model's design; adding steps mostly adds cost.
• Safety filters are baked in. BFL's models are the most safety-filtered in the field and the community's "lobotomized" complaint about FLUX 2 is documented. The node inherits that behavior - it's not something you can toggle off.
• Cost is per run but cheap per image. This is one of the least expensive cloud FLUX options, which is why it's the default suggestion for "try FLUX."

FLUX.2 [klein] 4B is the best "just works" image model in the pack. Low steps, low CFG, low price, strong prompt-following - a combination that's hard to beat for everyday generation.