arkennemasis fal · Flux 2 Klein 4B Base Edit LoRA · $0.016/MP

This is the base edit endpoint with a loras box bolted on: the un-distilled Klein 4B that edits a photo and, additionally, applies up to three LoRA weights from URLs. The price reflects it - $0.016 per megapixel against $0.009 for the same edit without a LoRA. Roughly double, for attaching weights.

Whether that's worth it depends entirely on what you're attaching. A realism or skin LoRA on an edit is a real change in output. A LoRA that was trained on the wrong size model is a very expensive way to get mush.

The LoRA box, and the trap in it

loras is a multi-line JSON box, not a file picker, and that's the first thing to grasp: the pack can't see your ComfyUI models/loras folder, because the weights have to be reachable by URL for fal to fetch them. In practice that means a direct download link from a model host. The field is fal's own list format with a maximum of three entries - check the model page for the exact shape fal expects - and leaving it empty is fine if you just want the plain edit.

The trap that costs people real money: 4B and 9B LoRAs are not interchangeable. People mixed them on CivitAI for weeks around launch, and the result isn't a subtle quality dip, it's a broken image. If the LoRA card says FLUX.2-klein-base-9B, it belongs on the 9B base nodes in this pack, not here. Size is one axis, base-vs-distilled is another; a LoRA trained on base 4B is at home on this node.

The rest of the inputs

prompt - command language, not scene description. remove blur and noise, fix exposure and color profile, micro-contrast. Name what must not change, or the model will invent a new subject for you.

negative_prompt and guidance_scale (5) - present because this is a base model with a real unconditional pass. The distilled nodes next door don't have these fields at all.

image_1 … image_6 - the first four are used, batches ride along each socket, order is instruction.

image_size - defaults to square_hd even though the tooltip says an unset size falls back to the input's dimensions. Set it deliberately if the frame matters. custom uses the width/height boxes.

num_inference_steps (28 - the community runs base at 50, and steps are free at per-megapixel billing), 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 with request id, estimate, saved files and fal's full response).

Install

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

Restart ComfyUI; it lands under arkennemasis → fal → Image → Flux. ComfyUI-Manager → Install via Git URL with the same repo also works. The only pinned dependency in the pack is httpx, for its ChatGPT/Codex nodes - the fal provider talks HTTP with the standard library.

Key in the .env next to run_nvidia_gpu.bat:

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

Where people get burned

Stacking without checking the size. Three 9B LoRAs on a 4B base endpoint is the classic expensive mistake. Klein does tolerate LoRA stacking better than Z-Image - the community ran three at full strength without burning - but only when they match the model.

A bad URL failing late. The weights have to download on fal's side, so a private or rate-limited link is a failed run you still waited for. Test with one LoRA before you build a stack.

Cost creeping up. You're now at $0.016/MP plus 1MP of input per edit, so a 1MP edit is about $0.032 and num_images multiplies it. Keep max_cost_usd low while testing - the check happens before upload, so a tripped cap is free. And if a rerun looks suspiciously instant and identical, that's reuse_identical_run serving the cached result rather than charging again.