Crop For Inpaint 🆎

Here's the dirty secret of whole-frame inpainting: you send the entire image through the sampler, and even the "untouched" pixels get degraded by the VAE encode/decode round trip. The crop-and-stitch pattern exists to fix exactly that - crop around the mask, inpaint only that region, paste it back so every pixel outside the fix is byte-for-byte identical. It's the pattern the knowledge base calls the structural answer to the oldest rule in inpainting, and Crop For Inpaint 🆎 is the crop half of it, paired with Stitch Inpaint 🆎.

The payoff is the thing everyone actually wants from inpainting: a 64px eye gets 1024px of generation budget because the crop renders at native resolution, and the rest of the image never passes through a model at all.

What it does

You feed it the image and a mask (white = what to repaint). It grows the mask's bounding box outward by context_factor (default 1.2) so the inpainter sees surrounding context, plus optional flat context_pixels if you want guaranteed padding. It then crops that region and emits:

• image - the crop, sized for the sampler.
• mask - a hard-edged sampling mask matching the crop. By design it is never feathered; feathering lives separately in the stitcher's blend mask, used only at paste time.
• stitcher - an AUSBOSS_STITCHER that carries the canvas, the rectangle, the blend mask, and the scale to Stitch Inpaint.

The stitcher is the clever bit: it broadcasts across a video batch, so one crop serves a whole clip. Crop once, run the inpainter over every frame, stitch once.

The inputs that matter

• blend_pixels (default 16) - feather width for pasting back. 0 pastes hard. This is not the sampling mask; don't confuse the two.
• output_multiple (default 8) - rounds crop and target up to a multiple so samplers accept them.
• target_width / target_height / target_megapixels - rescale the crop to a sampler-friendly area. 1.0 megapixel suits SDXL-class models; 0 keeps native size. Explicit target width/height overrides the megapixel target.
• mask_grow - dilates (or with negatives, shrinks) the sampling mask so the inpainter repaints past your drawn edge. Different job from blend_pixels.
• invert_mask - inpaint everything outside the drawn region.
• rescale_algorithm - the filter for the sampler round trip. bilinear is the safe default; nearest never invents pixels.

And the extend_left/right/up/down inputs grow the frame itself for outpainting - the new bands become part of the mask and of the stitched output, which makes this node do outpainting with the same bit-identical guarantee on the original pixels.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ausboss/ComfyUI-AusBoss.git

Restart ComfyUI and search for AusBoss (ComfyUI Manager: search ComfyUI-AusBoss). No extra Python dependencies for the core pair; fix_edge_halo on the stitch side wants the optional pymatting. Minimum ComfyUI 0.27.1.

Troubleshooting

An empty mask selects the whole image and stitches back unchanged - useful as a passthrough, confusing if you meant to inpaint something and the mask never arrived. If seams show despite a feathered blend, that's blend_pixels being too small or the classic double-blend rim - flip on Stitch Inpaint's fix_edge_halo. And if the crop comes back at the wrong size, check output_multiple isn't rounding it into a different dimension than your sampler expects.