Crop For Inpaint

Cuts out the part of a picture you masked, with some of what is around it, so the model repaints just that part at full detail instead of the whole picture shrunk down. Use it with Stitch Inpaint 🆎, which pastes the result back exactly where it came from. The Extend rows (extend_*) add new space on a side of the picture, which is how the pair outpaints.

Controls

image: BHWC image or video frames to crop around the mask.
mask: White marks the area to inpaint. A multi-frame mask is unioned into one crop window so a whole video shares one stitcher; an empty mask selects the full image (and stitches back unchanged). A picture with no mask painted on it stops the run with "No mask painted" and how to paint one (core Load Image sends a small blank stand-in in that case). A mask of another size than the picture also stops the run, and the message gives both sizes.
Context (context_factor): Grows the mask bounding box symmetrically by this factor. The grown window is shifted back inside the frame first. When it's bigger than the picture, Stay in picture decides what happens.
Stay in picture (keep_inside, on by default): Keeps the crop inside the picture. When the context would reach past an edge (a big mask, or a small picture), the crop stops at that edge, so the model sees only real picture and, after a Target rescale, paints the new part larger and sharper. Off is the old behavior: the crop runs past the edge and the extra space is filled with stretched copies of the edge pixels. One catch with it on: the model sees more of what's around your mask and carries it on, so paint over leftovers at the mask's edge that you don't want kept.
Blend (blend_pixels): Feather width used only when pasting back: the paste mask is widened by this many pixels and blurred. The sampling mask sent to the inpainter stays hard-edged, so feathering never weakens what the model sees. 0 pastes with the raw mask.
Multiple (output_multiple): The crop's size is rounded up to a multiple of this. Keep it at 32 (the default) for Qwen Image 2.1, Flux and most newer models. Their VAE trims a picture that is not a multiple of 16 or 32, and the trimmed result, stretched back over the crop, looks slightly bigger than the original. Stitch Inpaint says so in the console when that happens.
Target width / Target height (target_width / target_height, in the Target size, extend fold): Rescale the crop to a fixed size for the sampler. 0 keeps the native crop size; setting only one dimension derives the other from the crop's aspect ratio. Explicit values here override target_megapixels. When the crop is rescaled, the crop area takes in a few more pixels of context on one side so it has the exact shape of the target size, and the picture is scaled the same amount both ways, never stretched.

Shaping the mask

Grow (mask_grow): Dilate the sampling mask by this many pixels (negative shrinks) so the inpainter repaints past the drawn edge. This reshapes what gets painted, unlike blend_pixels, which only feathers the paste-back.
Blur (mask_blur): Gaussian sigma softening the sampling mask's edge, for models that honor soft masks. 0 keeps the hard edge.
Invert mask (invert_mask): Inpaint the black area instead of the white area — everything outside the drawn region. Applied before growing or blurring.

Sizing the crop

On the card, the Target size, extend fold holds these rows, Target width / Target height and the Extend rows.

Extra context (context_pixels): Flat extra context in pixels added around the mask box after context_factor's growth.
Target (target_megapixels): Rescale the crop for the sampler so its area is about this many megapixels — 1.0 suits SDXL-class models, 0 keeps the native crop size. Explicit target_width/target_height wins.
Rescale (rescale_algorithm): Resize filter for both directions of the sampler round trip. bilinear is the safe default, bicubic keeps upscales a touch crisper, area suits heavy downscales, nearest never invents pixels.

Outpainting

Extend left / Extend right / Extend up / Extend down (extend_left / extend_right / extend_up / extend_down): Grow the picture itself by this many pixels on that side. The new bands are filled with stretched copies of the edge pixels, are always painted (even with Invert mask on), and become part of the stitched output — draw nothing and the extended bands alone are inpainted.

Outputs

image: The cropped region, sized for the sampler.
mask: The matching hard-edged sampling mask (soft only where mask_blur says so).
stitcher: Everything Stitch Inpaint needs to put the result back.

Wiring

Load Image ── Crop For Inpaint ── image ──> LaMa Inpaint ── image ──> Stitch Inpaint
                   │  └── mask ─────────────────┘                          │
                   └── stitcher ───────────────────────────────────────────┘

Any inpainting sampler fits between the two nodes the same way — connect the crop's image and mask to it, then its output to Stitch Inpaint's inpainted input.

A picture with see-through parts keeps them through the pair. ComfyUI's Load Image splits them off into its mask; wire its IMAGE and MASK into Join Image with Alpha to put them back on before the crop. The sampler gets the colours, and Stitch Inpaint gives the result its clear parts back.

Technical details

Starting values: Context 1.2, Blend 16 px, Multiple 8; Grow, Blur, Extra context, the Target rows and the Extend rows start at 0.
With no Target size set, the crop is still rounded up to Multiple, so with Stay in picture on it can still reach a few pixels (at most Multiple - 1) past an edge.

6.0 KiB Raw Permalink Blame History