arkennemasis fal · Flux 2 Klein 4B LoRA · $0.01/MP

Klein's quiet superpower in the first half of 2026 wasn't speed, it was that it does styles. Flux.1 never could - the community comparison went straight to the point: "Klein does styles very well, that's something Flux.1 couldn't do, and it created images that astonished me." So a node that generates with up to three LoRAs attached is a more interesting proposition here than it would be on most checkpoints.

This is the distilled 4B with a loras box: four-step sampling, Apache 2.0 weights, $0.01 per megapixel. Twice the plain 4B node, which you're paying for the LoRAs rather than for better generation.

The LoRA box, honestly

It's a JSON text box, and it can't see your local files - fal downloads the weights, so what goes in the box is a URL-reachable weight file, in fal's own list format, up to three entries. Check the model page for the exact JSON shape. Leave it empty and you're just paying double for the plain model, so don't.

Two constraints that are hard, not soft. Klein 4B and 9B LoRAs do not interoperate - mixing them gives you mush, which is the most expensive kind of mistake on a paid endpoint. And many published Klein LoRAs target a base checkpoint and are then applied at inference to the distilled model, which works; if you're unsure which side a LoRA belongs on, the pack has both.

Stacking, though, is genuinely fine: Klein tolerates multiple LoRAs at full strength where Z-Image Turbo starts burning past two. Three is the ceiling here.

The inputs

prompt - sentences, not tags. The encoder is a Qwen3 LLM behind a chat template, and prompt weighting is disabled for it entirely, so (style:1.3) does nothing. What works is plain language with explicit ownership: "A realistic photo of a woman sitting on a beach. She is looking at the camera." And if your LoRA has a trigger word, use it - a silent LoRA is usually an uninvoked one.

num_inference_steps - 4 by default. Go to 8. The documented effect on distilled Klein is fewer mangled hands and cleaner text, and because fal bills per megapixel rather than per step, the extra steps are free.

seed - -1, or locked when you've found the take you like. Klein gives real variety between seeds, unlike some of the turbo-style models.

image_size - presets (landscape_4_3 default) or custom with image_size_width / image_size_height. num_images - the cost multiplier. output_format - png, jpeg, webp. enable_safety_checker - flagged images return black, and disabling needs account authorisation.

There's no guidance_scale and no negative_prompt here: distilled, so guidance is baked in and a negative prompt has no unconditional pass to attach to.

Outputs

images is the IMAGE batch - into SaveImage, an upscaler, a board for comparing variants. image_paths gives the saved locations as a STRING, timings is fal's timing JSON, has_nsfw_concepts reports the checker, seed is an INT, prompt echoes what the model saw, 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, and the node is under arkennemasis → fal → Image → Flux. ComfyUI-Manager → Install via Git URL with the same repo does it without a terminal. The pack's requirements.txt contains one line, httpx, and that's for the ChatGPT/Codex nodes - the fal side is standard library only.

Key, in the .env alongside run_nvidia_gpu.bat:

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

Where people get burned

Paying double for no LoRA. If the box is empty, use the cheaper non-LoRA node in this pack.

A LoRA URL that fails mid-run - private links, throttled hosts. Test one weight before you stack three.

Local-file confusion. This node has no idea what's in ComfyUI/models/loras. It only ever sees URLs.

Cached reruns. reuse_identical_run defaults on, so identical inputs return the stored files from output/fal instead of billing again - good for your wallet, surprising when you expected variation. max_cost_usd (default $20) is checked before upload, so an over-budget run costs nothing at all.