arkennemasis fal · Flux 2 Klein 9B Base Edit LoRA · $0.02/MP

The most expensive node in this Klein batch, and worth being clear about why: it's Klein 9B base - the undistilled model, with working guidance and a real negative prompt - doing instruction edits with up to three LoRA weights applied on top. $0.02 per megapixel of input and output, so a 1024×1024 edit is about $0.04. Compare the distilled 4B edit LoRA node at a quarter of that.

What justifies the price is the combination: base's controllability plus the LoRAs that were actually trained on 9B base. A consistency LoRA that fixes the family's identity drift, or a realism LoRA that takes the plastic sheen off skin, is doing something the plain nodes can't. A style LoRA that just changes the look is usually cheaper to run on a smaller endpoint.

The LoRA box

loras is a JSON text box, and the weights have to be reachable by URL - fal downloads them, so your ComfyUI/models/loras folder is invisible to this node. It takes fal's list format, up to three entries; the model page has the exact shape. Leave it empty and you're paying 9B-base-edit prices for nothing extra.

The constraint that costs people real money: 9B LoRAs don't work on 4B endpoints and vice versa. Klein's own release split everyone's files across two sizes and the first weeks of CivitAI uploads were a mess of people mixing them. The other axis is base versus distilled - and here you're on the right side of it, because most published Klein LoRAs train against FLUX.2-klein-base-9B, which is exactly this checkpoint.

One more: read the LoRA card for a trigger word. A LoRA that seems to do nothing is usually one your prompt never invoked.

The inputs

prompt - a post-processing command list. remove blur and noise, fix exposure and color profile, micro-contrast. Name what must survive the edit.

negative_prompt and guidance_scale (5) - live, because base runs a real unconditional pass. image_1 required, plus image_2–image_6: first four used, batches per socket, socket order is reference order. image_size - square_hd preset by default; the tooltip says an unset size uses the input's dimensions, so set it deliberately if the frame matters. num_inference_steps 28 (base wants more like 50, and steps are free at per-megapixel pricing), acceleration, seed, num_images, output_format, enable_safety_checker.

Outputs

images - the edited IMAGE batch, image_paths (STRING), timings, has_nsfw_concepts, seed (INT), prompt (STRING), and info (JSON: request id, cost estimate, saved files, fal's full answer).

Install

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

Restart ComfyUI, then find it under arkennemasis → fal → Image → Flux. ComfyUI-Manager → Install via Git URL with the same repo is equivalent. The only pinned dependency in the pack is httpx (for the ChatGPT/Codex nodes); the fal provider is standard-library only.

Key, in the .env by run_nvidia_gpu.bat:

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

Where people get burned

A dead LoRA URL. fal has to fetch the weights, so private or rate-limited links fail a run you waited for. Test one before stacking three.

Cost blindness. At $0.02/MP plus a megapixel of input, three variants of a 2MP edit is 12 cents - cheap, but only if you meant it. num_images is the multiplier to watch.

The warm shift and the drift. Edits come back slightly warm versus the source, and identity still isn't held the way Qwen-Image-Edit holds it. A consistency LoRA narrows the second problem; nothing fixes the first except a colour-match.

Reruns served from cache. reuse_identical_run defaults on so identical inputs aren't billed twice - turn it off for a different take. max_cost_usd (default $20, evaluated before upload) is the leash that makes experimenting cheap; a run that trips it never leaves your machine.