flux2Klein / createImage

Flux 2 Klein is the small-model line Black Forest Labs shipped to fix Flux 2's "nobody can run it" problem - and it's genuinely small, which means this cloud node is the cheap way into the Flux 2 family. Where the Flux 2 Dev recipe is a heavyweight job, Klein runs four checkpoints - 4B and 9B, each in distilled and base form - and this node lets you pick which one via a model_version dropdown. No model files land on your machine; the sdcpp fleet does the work and bills you in Buzz.

The pick that matters is 4b vs 9b. The 4B is Apache 2.0, the commercially-usable one, and community testing found it genuinely underrated - fast and capable, especially for a first pass. The 9B carries the FLUX non-commercial license and is the better-looking daily driver. Distilled (just "4b"/"9b") is what you generate with; the -base variants are training foundations that BFL itself says aren't meant for generation quality. If this is your first Klein render, start with 4b - it's the one you can actually do something with.

How it works

Standard pack client: submit engine: sdcpp, ecosystem: flux2Klein, operation: createImage with modelVersion selecting the checkpoint, long-poll, get a ComfyUI IMAGE back. Note the category here is just Civitai/Image/flux2Klein - the engine is still sdcpp, the menu just doesn't show it because Klein isn't reachable through another engine.

The inputs that matter

• prompt (required) - Klein's Qwen3-4B/8B encoder wants natural language; write sentences.
• model_version - 4b, 4b-base, 9b, or 9b-base. Default 4b. Don't mix LoRAs across the 4B/9B lines - they're not interchangeable.
• steps / cfg_scale - 20 / 5 defaults. The distilled variants are trained for very few steps (4-ish locally), but this recipe defaults to 20; dropping steps is where you'll actually save Buzz.
• width / height - 1024 default, 512–2048 range.
• negative_prompt - exists; Klein, like the rest of Flux 2, barely needs one.
• loras - via a Civitai LoRA Selector; most Klein LoRAs target 9b-base and are applied at inference.

Install

Pack install: ComfyUI Manager → search "Civitai Comfy Nodes" (publisher civitai) → Install → restart. Or git clone https://github.com/civitai/civitai-comfy-nodes.git into custom_nodes, pip install -r civitai-comfy-nodes/requirements.txt (just requests), restart.

Common issues

Credentials first: Civitai Auth node into api_config or the CIVITAI_API_TOKEN env var (from civitai.com/user/account); headless installs need the env var.

The Klein-specific mistakes are version confusion and LoRA mixing. Picking 9b-base because "base must be better" gives you the training foundation, not the good generator - that's the most common disappointment. And applying a 4B LoRA to a 9B run (or vice versa) silently degrades quality rather than erroring.

And because each generation costs Buzz, remember the distilled variants are the whole point: if you're just iterating, low steps on 4b is the cheapest path to a decent result.