arkennemasis fal · Flux 2 Klein 4B Base · $0.009/MP

Klein shipped as four checkpoints, and the one thing the internet got wrong for weeks is that "base" isn't a bigger or better version of "distilled" - it's the un-distilled one. Base means the model still needs real classifier-free guidance and 50-ish steps to do its job. Its whole reason to exist is training: it's the target that most published Klein LoRAs are trained against.

So this node has an honest use and a dishonest one. The dishonest one is treating it as a better Klein 4B for generation - Black Forest Labs say it plainly, base "is not intended for image generation as the quality isn't as good as the distilled normal model." The honest one is running a base-tuned LoRA at inference, or checking what a base fine-tune does before you commit an overnight training run.

The inputs

prompt - sentence-style, because the encoder is a Qwen3 LLM reading a chat turn. Tag weighting does nothing here.

negative_prompt - note this field exists on the base node and not on the distilled 4B sibling. That's the distillation story in one line of UI: base runs a real unconditional pass, so a negative prompt has something to push against. Use it.

guidance_scale - 5 by default. The community number for base Klein is 4-5, so fal's default is in the sane band; unlike the distilled variants there is no "keep it near 1" rule here.

num_inference_steps - 28 on the node, against the 50 steps people actually run base Klein at. Fal bills per megapixel, so extra steps cost seconds and not cents.

acceleration - none / regular / high, defaulting to regular. This is the base-and-Dev flavour of the family; the distilled nodes don't have it.

image_size - presets with landscape_4_3 as the default, or custom plus image_size_width / image_size_height. One megapixel is the natural point.

num_images, output_format, enable_safety_checker - the last one flags rather than errors: a refused render comes back black, and switching it off needs authorisation on your fal account.

Outputs

images (IMAGE, batch if num_images > 1), image_paths (the saved files as a STRING), timings, has_nsfw_concepts, seed (INT), prompt (STRING), and info - JSON with the request id, cost estimate, saved files and fal's raw answer.

Install

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

Restart, then look under arkennemasis → fal → Image → Flux. ComfyUI-Manager → Install via Git URL with the same repo works too. The pack declares exactly one dependency, httpx, and that's for its ChatGPT/Codex nodes; fal runs on the standard library, so this node adds nothing to your environment.

Key in the .env next to run_nvidia_gpu.bat:

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

Where people get burned

Comparing base and distilled at the same settings. They're different models with different habits. If you run base at 4 steps because that's what the distilled node defaults to, you'll conclude base is broken. It isn't; it just hasn't been taught to skip.

Thinking it's the licensed one. Both 4B checkpoints are Apache 2.0 - the commercially usable tier of Flux 2. That's why 4B base is the smart foundation if you ever intend to train something you can ship, which is a road almost nobody has taken.

Forgetting this is still a 4B model. Base fixes the distillation artefacts, not the ceiling: detail and prompt comprehension are 9B territory, and the anatomy reputation is shared across the line.

Paying twice for the same render. reuse_identical_run defaults on, so identical inputs with the files still in output/fal are returned rather than re-billed. Flip it off when you want a fresh take. And if a run halts before uploading, that's max_cost_usd (default $20) doing its job before the paid call.