Inpaint

Inpainting in stock ComfyUI is a small chore: you wire a VAE encode, a patch, a sampler, a decode, and a composite, and you hope the pieces line up. LF_Inpaint collapses that whole chain into one node. You hand it an image, a mask, a model, a CLIP and a VAE - it runs the diffusion pass and returns the inpainted image. If you've ever stared at a "paint out that power line / fix that face / remove that watermark" workflow and wished it were a single box, this is that box.

It sits squarely in the "mask-based inpainting" camp, and the knowledge base has a clear-headed take on why that still matters in the era of instruction-edit models: an edit model regenerates everything and drifts, while a mask inpainter leaves the unmasked pixels bit-identical. If you need exactly-one-region changed and nothing else moved, this is the tool.

How it works

The node runs the diffusion itself - model, clip, vae go straight in, no separate KSampler, no latent encode/decode nodes. Required basics: mask (1 = inpaint, per-image masks/batches supported), steps (default 16), cfg (default 7), and the critical one, denoise_percentage. It's 0–100 rather than the 0–1 denoise you usually see: 40% (the default) means a moderate rewrite of the masked area, ~10–25% for subtle fixes, 80%+ for full regeneration inside the mask.

Two ways to drive the generation: either positive_prompt / negative_prompt strings (simplest), or pre-built positive_conditioning / negative_conditioning from a CLIP Text Encode node if you want conditioning the rest of your graph already built. You can also swap sampler (44 options, default dpmpp_2m) and scheduler (default karras).

The gotcha worth knowing

upsample_target defaults to 1024. If your input is smaller, the image and mask get upscaled to ~1024 on the long side before inpainting and pasted back down afterward. That's actually the intended detail-boosting behavior - small crops get a bigger generation budget - but it's easy to mistake for the node resizing your image. Set it to 0 if you want no upsampling at all.

Outputs

image / image_list - the inpainted result. Compare widget shows before/after. That's the entire interface.

Installing it

lf-nodes installs as one pack:

cd ComfyUI/custom_nodes
git clone https://github.com/lucafoscili/lf-nodes

then restart ComfyUI (or ComfyUI Manager → "LF Nodes"). No extra model - it runs whatever checkpoint your graph already loads.

Troubleshooting

• Too much change - the 40% default is a rewrite. For a stray speck or a small flaw, drop denoise_percentage to 10–20 first.
• Seams at the mask edge - feather the mask before it reaches this node (or use the advanced variant, which feathers internally).
• Squashed output - almost certainly upsample_target doing its job on a small input. Set it to 0 and try again.