arkennemasis fal · Flux 2 Klein 9B Edit · $0.011/MP

FLUX.2 [klein] is what most people actually use to edit a photo in 2026, and the 9B distilled variant wants something like 20GB of VRAM before you've loaded an upscaler. This node is that exact model - same weights, same 4-step distillation - running on fal's GPUs, billed per megapixel. If you have the card, run it locally. If you're on a 12GB 3060 and just want to change a jacket colour, $0.011 per megapixel is a rounding error next to the render time.

What it really is

It's a thin wrapper, not a reimplementation. One node per fal endpoint, generated from a JSON file in the pack's fal_provider/models/ folder - this one points at fal-ai/flux-2/klein/9b/edit. When you hit Run, the node uploads your picture to fal's CDN, submits a job to fal's queue, polls it, and downloads the result into output/fal/fal-ai_flux-2_klein_9b_edit/. From the canvas it looks like a sampler. Underneath it's an HTTP client. Your input image leaves your machine - that's the mechanism, not a bug - and klein here is the distilled 9B: 4 steps, guidance baked in, a Qwen3-8B encoder behind the scenes.

The inputs that matter

prompt is an edit instruction, not a scene description. The failure everyone hits is the model dropping into "generator" mode and inventing a new person. Prompt the file instead: "remove blur and noise", "fix exposure and colour profile", "zoom out until her white sneakers are visible. Maintain the faces. Maintain the posture." It reads like a Photoshop command list and it works much better than adjectives.

image_1 is the picture to edit. There are extra sockets up to image_6, but fal takes a maximum of four images total, and a batch on one socket counts as several - they're numbered in socket order, first picture of socket 1 is number 1.

num_inference_steps runs 4 to 8, default 4. Take the 8. The well-known Klein hack - 8 steps at slightly higher guidance - is the thing that kills most of the extra-fingers problem, and steps here are cheap.

image_size defaults to square_hd, which is 1024×1024. The tooltip says it uses the input size if not provided, but there's no auto option in this node's list, so a 3:4 portrait comes back square unless you switch to a portrait preset or custom and set the width and height yourself.

num_images (1–4) multiplies your bill. enable_safety_checker is on by default; flagged images come back as black images, and turning it off requires authorisation on your fal account, so unauthorised requests are checked regardless.

Four options sit on every fal node: max_cost_usd (default $20, 0 disables) stops the run before anything is uploaded if the estimate is over; reuse_identical_run returns an already-paid result for identical inputs; extra_json merges a raw object into the request for fields with no box; max_concurrent (default 1) limits how many paid fal calls run at once across the whole graph.

Outputs

images wires into anything expecting an IMAGE - SaveImage, your upscale chain, a mask node. image_paths gives you the saved file locations as text. timings, has_nsfw_concepts, seed and prompt are fal's own scalar returns, and info is a JSON blob with the request id, cost estimate, saved files and fal's full answer - the honest record of what you were charged.

Install

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

Restart, then put one line in the .env next to run_nvidia_gpu.bat (portable) or in your ComfyUI folder:

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

There is no key field on the node, deliberately - a typed key ends up saved inside every workflow you share. The fal provider uses nothing but Python's standard library, so there is no pip step for these nodes; the httpx in requirements.txt is for the pack's Codex/ChatGPT nodes.

Where people get burned

An all-black image is the safety checker, not a crash. A warm colour shift on small edits is normal for Klein. Identity drift is the model's known weakness - pair it with a consistency LoRA, or reach for Qwen-Image-Edit if the face must survive. And if a run reports "same inputs as request … reused, nothing billed" when you wanted a fresh take, that's reuse_identical_run doing its job; turn it off or change the seed. At 1MP out and 1MP in you're paying $0.022 a shot, so four images and four tries is real money but not painful.

One last thing worth knowing: nobody on Reddit mentions this pack by name - an exact-phrase search of the corpus returns zero threads. There's no community folklore to inherit, so the README and the node's tooltips are the documentation.