arkennemasis fal · Flux 2 Klein 4B · $0.005/MP

Klein is what happened after Black Forest Labs shipped a 32B model nobody could run and then fixed it properly. "Klein" is German for small, and the 4B distilled variant is the interesting one: Apache 2.0 weights, four-step sampling, runs on an 8GB card locally, and it does text-to-image and editing in the same model. On fal this node drives it for $0.005 per megapixel - half a cent for a 1024×1024 image. That's cheap enough that batching four to pick one is the sane default.

It's also, per the knowledge base, "very slept on" - people assume small means weak, then find it handling virtual try-on with three references at close to a 100% hit rate.

What you're actually running

The distilled 4B: 4 inference steps, and guidance baked into the weights instead of applied at sample time. That second part is why this node has no guidance_scale field and no negative_prompt field - that's not the pack dropping inputs, that's the model. There's no unconditional pass for a negative prompt to steer, so the box would be a lie.

The price of that shortcut is anatomy. Extra fingers and the occasional third limb are the documented, unfixed complaint about Klein, and on four steps you'll hit them at volume. The community's fix, which works, is worth stealing: run 8 steps instead of 4. The launch-week report was blunt - 8 steps with guidance nudged slightly above 1 fixes "body horror", finger counts and text. There's no guidance field here, but num_inference_steps is right there, and fal bills per megapixel, not per step. Two extra runs' worth of wall time, zero extra dollars.

The inputs

prompt - Klein's encoder is a Qwen3 LLM wrapped in a chat template. It reads sentences, not tags. Prompt weighting is disabled for this text encoder, so ((face:1.4)) does exactly nothing; what works is plain language with clear ownership: "A realistic photo of a woman sitting on a beach. She is looking at the camera." Structured JSON-ish prompts reportedly do well too, because you're talking to an LLM.

num_inference_steps - 4 by default. Try 8.

image_size - landscape_4_3 by default, or custom with image_size_width / image_size_height.

num_images - the cost multiplier, and at this price the right way to work. Four 1MP renders is two cents.

output_format - png, jpeg or webp.

enable_safety_checker - on. A flagged render comes back as a black square; disabling it requires account authorization on fal, and unauthorised requests get checked anyway.

Outputs

images is the IMAGE output (batched if you asked for several) - into SaveImage, an upscaler, or a contact sheet. image_paths is where they landed as a STRING, timings is fal's timing JSON, has_nsfw_concepts tells you what the checker thought, seed is an INT to lock a good one, prompt echoes what the model received, and info is JSON with the request id, cost estimate, saved files and fal's full answer.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart ComfyUI and it's under arkennemasis → fal → Image → Flux. ComfyUI-Manager → Install via Git URL also works. The pack's only declared dependency is httpx, for its ChatGPT/Codex nodes; the fal provider uses Python's standard library, so the install is the clone and nothing else.

Key, one line in the .env by run_nvidia_gpu.bat:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

Where people get burned

Anatomy, again. No prompt fixes it - the failure is baked in during the first couple of steps, which is why the mitigation is steps and batching rather than wording. Batch four, pick the best, inpaint the hand.

Signature hallucination. Klein base likes to drop a mark in the bottom-right of stylised images, and with no negative prompt at CFG 1 there's nothing to argue with. Crop or inpaint if it lands somewhere ugly.

Cached reruns. reuse_identical_run is on by default, so the same settings plus the same files in output/fal returns the previous result without charging. Turn it off for variation. And if a job dies after leaving fal's side, the pack's free Recover Result node collects it by request id from output/fal/_requests.jsonl instead of you paying twice.