arkennemasis fal · Flux General Inpainting · $0.075/MP

Instruction editing ate most of inpainting's lunch between 2025 and 2026. Ask a model to change the jacket and it changes the jacket, no mask required. What instruction models can't do is leave the rest of the frame alone - they take the whole picture into context and emit a whole new one, so unmasked regions come back close but not identical, and the drift compounds if you chain edits.

That's the niche this node occupies: it's the mask path, for when you want to say where and not just what. It runs the Flux General endpoint, so it also carries the LoRA/ControlNet/IP-Adapter extensions and the NAG negative-prompt controls.

The three inputs that matter

image - required, the picture being edited.

mask - required, and it's a proper ComfyUI MASK, not a URL. Draw it with the mask editor or wire a segmentation node into it. The pack uploads it as a PNG sized to match your first image, so you don't have to think about resolution - but you do need the mask to line up with the image you actually connected, because a mask from a different crop is the single most common way to get a nonsense result.

strength - defaults to 0.85: 1.0 completely remakes the image, 0.0 preserves the original. This is the dial people get wrong. The classic inpainting guidance is that 0.3–0.6 handles most real work - swapping a garment, changing an expression - and going to 0.8+ is essentially text-to-image inside the mask, composition included. The default here is at the aggressive end, so lower it for detail work.

The rest of the node

Because it's the general pipeline: prompt and negative_prompt, with negatives routed through NAG - nag_scale (3, don't exceed it), nag_tau (2.5), nag_alpha (0.25), nag_end (0.25). The JSON extension fields - loras, control_loras, controlnets, controlnet_unions, ip_adapters, easycontrols - take URLs fal can fetch, not local paths. use_real_cfg (false) costs time and money when enabled. Sampling: image_size (square_hd default, custom for exact dimensions - watch this, since a mask and an image at different aspect ratios will not line up), num_inference_steps 28, seed -1, guidance_scale 3.5, scheduler euler, base_shift 0.5, max_shift 1.15, use_beta_schedule false.

There are also reference_image (with reference_strength, reference_start, reference_end) and six optional fill_image_* sockets for reference-only and Redux-style fill prompting.

Outputs

images and image_paths, saved under output/fal/fal-ai_flux-general_inpainting/, plus timings, has_nsfw_concepts, prompt, seed, and info carrying the request id, cost estimate, saved files and fal's answer.

Install

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

Restart ComfyUI, or use ComfyUI-Manager → Install via Git URL. requirements.txt declares a single dependency, httpx, and it's for the pack's Codex nodes; the fal nodes run on the standard library and download no weights.

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

goes in the .env next to run_nvidia_gpu.bat (portable build) or in your ComfyUI folder. No fal node in this pack has a key field - the design decision being that a key in a widget ends up inside shared workflow files.

The caveat, said plainly

You are not compositing locally. The endpoint generates a full frame with your mask as guidance, so pixels outside the mask are regenerated too - they'll look right, they just won't be identical. If your requirement is that nothing outside the mask may change, provably, then the hosted inpainting endpoints are the wrong tool and the local route is the right one: mask locally, run the model, composite the masked region back over the original. That gives you bit-identical untouched pixels, which no API round trip can.

Where this node genuinely wins: you want a mask-guided edit on Flux General without a 12B checkpoint and a ControlNet on your disk, or you want the extensions available at the same time. Then 7.5¢ a megapixel is a reasonable toll.

Cost

$0.075/MP - 1024×1024 ≈ 7.5¢, and use_real_cfg pushes it higher for the same output. max_cost_usd (default $20, 0 for no cap) stops the run before upload if the estimate is over. reuse_identical_run (default on) means the same image bytes, same mask and same settings return the file you already paid for rather than a fresh charge - useful, and worth knowing before you queue a re-run expecting a different sample.

When it goes wrong

The mask landed on the wrong thing. Check that the mask and the image are the same crop and the same orientation. The pack sizes the mask PNG to your image, but it can't know if that mask belongs to a different photo.

A seam appears at the mask edge. Lower strength first. The other classic fix is feathering the mask so the transition isn't a hard line - how much feathering depends on the size of the edit, and it's the field you'll be adjusting in your own mask pipeline rather than on this node.

A tiny detail came out mushy. That's resolution, not the model: the whole frame gets the generation budget. Crop the region, render it large, and paste it back if the detail is what matters.