arkennemasis fal · Flux Krea LoRA Inpainting · $0.035/MP

Every other node in this family regenerates the whole frame and hopes the untouched parts come back looking the same. This one takes a mask and is only supposed to change what's inside it. $0.035 per megapixel, hosted, LoRA-capable.

That property - the unmasked pixels stay put - is the reason masked inpainting still has a job in 2026, when instruction editors handle most "change this to that" requests better and without a mask. An edit model reads the whole image into context and emits a whole new one, so everything drifts a little, and it compounds across a chain. A mask doesn't drift. If you're compositing into a frame where the surroundings must survive, this is the tool.

Inputs

The mask pair is the whole point:

• image - the picture to edit. A IMAGE socket, one picture.
• mask - a ComfyUI MASK. The pack does the translation for you: ComfyUI masks are 1 where they're set, and this uploads an RGBA PNG that's transparent where you painted. It also resizes the mask to your image's dimensions, so a mask drawn at another resolution still lines up.
• strength, default 0.85 (0.01–1). 1.0 remakes the region outright, low values barely move it. This is the dial that makes inpainting worth doing by hand - "a bit darker" is a number here, not a sentence.

Then the usual generative settings: prompt (describe only what should appear in the hole - describing the rest of the scene just gives the model ideas), num_inference_steps default 28 (1–50), guidance_scale default 3.5 (0–35), seed, num_images 1–4, enable_safety_checker, output_format, acceleration.

loras is a multiline JSON string: LoRA entries with weights and scale, merged together. Empty leaves it out. They load on fal's side, so it's URLs - not your local models/loras folder.

image_size plus image_size_width / image_size_height (16–14142, step 16) are also on the node, defaulting to square_hd. Worth a glance: a size that disagrees with your source framing is a quiet way to get a subtly recomposed result back.

Below the model's own fields: max_cost_usd (default 20, checked before upload, 0 = off), reuse_identical_run (on by default - identical inputs and settings return the already-paid files), extra_json, and max_concurrent (default 1 across all fal nodes in a run).

Outputs

images for preview/save, image_paths for the files under output/fal/fal-ai_flux-krea-lora_inpainting/, then timings, seed, prompt, has_nsfw_concepts and info - the last of which carries the request id, the cost estimate and fal's raw answer.

Install

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

ComfyUI-Manager users can search comfyui-arkennemasis and let it handle the install. Nothing exotic in requirements.txt - httpx for the pack's unrelated Codex nodes; the fal provider runs on the standard library. Restart ComfyUI; the node appears under arkennemasis/fal/Image/Flux.

Then the key. One line in the .env in the folder with run_nvidia_gpu.bat:

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

No key widget on the node - that's deliberate, so a key can't ride along in a shared workflow.

Things that bite

Mask inverted in your head. Paint where you want the change, same as any ComfyUI inpaint node; the conversion to "transparent = edit here" happens inside the node. If your output changed everything except the region you painted, you've almost certainly wired a mask from an upstream node that produced the inverse.

Mask doesn't match the image. It gets resized to the image's dimensions, so it will line up - but a mask built at a wildly different aspect ratio gets stretched and your boundary moves.

Inpainting is billed on the whole image's megapixels, not the masked area, so a tiny touch-up on a big canvas costs like a big generation. The badge on the node is honest about it.

Black output. That's the safety checker, not a bug. has_nsfw_concepts will confirm it. Turning the checker off requires fal account authorization.