ausboss-ComfyUI-AusBoss/js/docs/AUSBOSS_NODES_StitchInpaint.md

Stitch Inpaint

Puts what the model painted back into your original picture, so everything you did not ask it to change stays exactly as it was. Wire in the stitcher from Crop For Inpaint 🆎, Load Image + Pad 🆎 or any Crop + Rotate + Pad 🆎 node. If the sampler changed the size of the result, it is resized to fit first, then joined to your picture the way Seam says.

Paste onto: crop or original

Paste onto picks what you get back. It lives in the gear menu, next to Seam.

• crop (the default) gives back the picture you framed.
• original puts the edit on your whole picture, at its full size. Use it after you crop by hand in Image Crop + Rotate + Pad 🆎: the model works on the crop, and everything you did not paint stays exactly as it was. It also works when you turned or resized the crop.

Padding has no place on your picture, so anything painted in added space is left out. With Crop For Inpaint and Load Image + Pad the result is your whole picture either way. The video nodes don't have this yet: their result stays on the crop, and the console says so once.

Seam: classic or blend in

Seam picks how the new area joins your picture. It lives in the gear menu: click the small gear in the card's top-right corner.

• classic (the default) is the paste every earlier workflow uses: the new area fades in over the feathered mask, and Tone match can shift its colour toward your picture. Saved workflows keep it, so they render exactly as before.
• blend in is for outpaints, turned or straight. The model redraws a thin strip along the edge of your picture, and its version never matches yours exactly. Blend in fades from the model's picture to yours a little way inside your picture, where the two already line up. With Tone match on, it first checks how the model changed the colours of that strip and takes the same change back off the new area, so a model that paints a little lighter, darker or warmer still meets your picture in tone. With Tone match at 0 the new area stays exactly as the model painted it. Fix edge halo is for classic only: under blend in it stays in place but dims, and a blend in chip shows next to the gear.

Blend in needs to know where your picture's edge is. Load Image + Pad and the Crop + Rotate + Pad nodes record it; Crop For Inpaint does not, so a Crop For Inpaint stitcher always pastes classic for now (the console says so once).

What blend in fixes: the lighter or darker band classic's tone match can lay over a turned outpaint, the grey haze a flat tone lift lays over a dark new area, and the smeared strip where the two pictures were cross-faded. What it can't fix: a pattern the model drew out of step at the edge, such as plaid or stripes running along it, or fill the model left unpainted. Those are in the model's own picture.

Guarantees

• With the classic seam, pixels outside the blend region are bit-identical to the original image — they never pass through a resize or blend. This holds with fix_edge_halo on as well: the toggle changes the color that is pasted, never how far the paste reaches.
• With blend in, your picture is bit-identical from where the blend ends, a few dozen pixels inside its edge (about 35 px at common settings). At Tone match 0 the new area is bit-identical to what the model painted, and the whole stitch is the 2.4.0 blend in, pixel for pixel.
• With the classic seam, feeding the crop back unchanged reproduces the original image exactly (with fix_edge_halo off; the fix rewrites the feathered band on purpose). Blend in does not promise this: it keeps the model's version of the thin strip at your picture's edge.
• A stitcher built from one image broadcasts over any number of inpainted frames. A stitcher built from more frames than came back is trimmed to the leading ones with a console note - video models keep 8n+1 (LTX) or 4n+1 (Wan) frames and drop the tail - so a long run is never thrown away over the count. More inpainted frames than source frames is still an error.

Pictures with see-through parts

A picture with a clear background, such as a cut-out, keeps it. Most models paint colour only, so the new area takes your picture's own see-through parts where it lands: what was clear stays clear, and the edit shows where your picture is solid. A model that paints see-through parts of its own (Qwen Image 2.1 can) has them blended in like the colours.

So with a colour-only model, new things painted into a clear area stay hidden. Fill that area in first if the edit needs it.

Controls

• stitcher: The stitcher from one of the nodes above. It carries the paste window, the blend mask, and the untouched original pixels.
• inpainted: The inpainted crop. A stitcher built from a single image broadcasts across an N-frame inpainted batch, so one still-image crop can stitch a whole video; matched N-to-N batches also work.
• Seam (seam, in the gear menu): classic (the default) or blend in. Read "Seam: classic or blend in" above. It has no socket, so a link cannot drive it.
• Paste onto (paste_onto, in the gear menu): crop (the default) or original. Read "Paste onto: crop or original" above. It has no socket.
• Fix edge halo (fix_edge_halo, classic seam only): Off by default. Recovers the true color under the feathered seam before pasting, so half-transparent edge pixels stop blending their background in a second time. It costs real time per frame — read "What it costs" below before turning it on for a whole batch.
• Tone match (color_match): 0 (off) to 1. Matches the new area's tone to your picture. Classic reads the drift across each seam; blend in reads it where the model repainted the edge of your picture. Read "Matching the tone" below.

Matching the tone

An outpaint often comes back a little off the picture it extends. Klein can darken the mid-tones of a night scene and brighten a daylight one, Krea 2 paints a touch darker, and Qwen Image 2.1 usually keeps your colours. Tone match takes that off, in either seam.

With blend in

Blend in compares the model's version of the strip along your picture's edge with your picture, pixel for pixel, and works out how the model changed each brightness level. It takes the same change back off the new area, so dark stays dark and a mid-tone the model darkened comes back up. It checks that the change holds up along the edge first (it tests the fit on part of the strip it did not use), and changes nothing when it does not, when the model kept your colours, or when there is no feather to read.

A wider feather gives it more to read. At feather 24 or more it removes most of a model's drift; at 10 to 12 it removes part of a strong one, and never more than the drift. A video clip is measured once for all its frames, so the correction cannot flicker.

With classic

An outpaint comes back a few percent off the picture it extends — Krea 2 a touch darker, Klein a touch lighter — and the padding colour makes no difference to that. The feathered sampler mask then smears the step into the band inside the source, so without a match the seam shows as a faded band even though the content continues perfectly.

color_match reads the drift across each padded side's seam, line by line: for a left or right side, every row compares the 24 generated pixels just outside the seam with the 24 original pixels just inside it; a top or bottom side does the same per column. Sky and water on the same seam drift by different amounts, and one number for the whole picture leaves one of them showing. Each line's reading is clamped to a few LAB units around its side's average — a pier post leaving the frame is content, not drift — and smoothed so no single line prints a stripe. Where two padded sides meet, their curves blend by distance, so the correction turns the corner without a crease.

The shift is subtracted from the pasted region before the blend. In the padding that is the whole shift; inside the feathered band the paste alpha scales it, which is exactly the share of each band pixel the sampler painted. Source pixels stay bit-identical. A crop stitcher, which has no source rectangle, gets a single global shift measured in its blend band instead.

On a Crop + Rotate + Pad canvas the source rectangle is the crop, so it can also hold the rotation's empty corners and any transparency. The inside pixels skip those, and a line with no picture inside its seam takes the global shift, read in the blend band along the picture's real edge.

The estimate assumes neighbouring strips depict similar content. It cannot reliably distinguish a tone shift from a different object or shadow at the seam, even with the clamp and smoothing.

For video, start with Tone match at 0. The estimate runs independently on each frame, so moving subjects or camera motion can turn those content differences into flickering dark or light bands across the generated area. Compare the decoded frames before stitching with the stitched result; if the bands appear only after stitching, disable color matching. The source paste and feathered blend still work with it off. Blend in's Tone match is measured once for the whole clip and every frame is blended with the same map, so blend in adds no flicker of its own.

Fixing an edge halo

This is for the classic seam. Blend in does not fade the new area in over the blend mask, so no background gets counted twice, and the toggle dims while blend in is on.

A feathered paste mixes each seam pixel with the original image. When the inpainted pixel is itself already a mix of new content and the old background — common after object removal, or when the mask hugged the subject too tightly — that background gets counted twice and the seam reads as a dark or light rim around the repair.

Turn fix_edge_halo on and the pasted color is re-derived first: the opaque core's color is spread outward across the blend band, and the paste blends toward that instead of the contaminated pixel. Leave it off when the seam already looks clean — a clean seam has nothing to gain, and the estimate is not free.

The fix needs the optional pymatting package, installed with the Python that runs ComfyUI:

python -m pip install "pymatting>=1.1"

Without it the node prints one console note and pastes exactly as it would with the toggle off — the graph keeps running either way.

What it costs

The estimate is a CPU solve run once per frame, and it scales with the paste window rather than with the whole image — roughly 90 ms per megapixel of that window. Measured on a 16-thread desktop CPU with pymatting 1.1.15: about 48 ms for the 768x768 window a 1024x1024 frame produces, about the same for the 960x544 window from 720p, and about 106 ms for the 1440x816 window from 1080p. Expect several times that on a modest laptop CPU.

Per frame, that adds up: 300 frames of 1080p is over half a minute spent in the estimate alone. This is why the toggle ships off — it is meant for finishing a take you have already chosen, not for a long exploratory batch. Batches of more than one frame report per-frame progress and check for a cancel between frames, so a run started by mistake stops at the next frame boundary rather than playing out to the end.

When the new area comes back empty

An outpaint model sometimes leaves part of the new area alone, and the result shows plain bars in the fill color. Stitch Inpaint checks every outpaint for that. When a quarter or more of the new area still shows the plain fill (or half of it in any checked frame of a video), a message is written to the console, without a popup. The picture is stitched as usual either way. This check can flag intentional results. If the image looks right, you can ignore this warning.

Try another seed, or describe the whole wider scene in the prompt. Adding less space at a time also helps. For a video, a dark clip paints better if you brighten it first.

Outputs

• image: The original-size image with the inpainted region blended in. After a crop in a Crop + Rotate + Pad node, that means the crop's size, unless Paste onto is on original.
• blend_mask: The feathered paste mask in the stitched image's own coordinates — white where the inpaint blended in, zero where the original survived untouched. Wire it into a color-match or compositing node to treat exactly the pasted region without rebuilding the mask by hand. It is batched like image, so a single-image stitcher broadcasting across a frame batch hands every frame its mask. With Seam on blend in it shows the blend instead: white on the new area, fading to black where your picture keeps its own pixels. It is still zero exactly where the stitch left the original alone.

Technical details

The rest of this page is for people who want to know how blend in works and how it was tested.

How Paste onto original works

Image Crop + Rotate + Pad's stitcher carries your picture and the crop's position. A straight crop at its own size is a plain paste, so every pixel the mask does not touch is bit-identical. A turned or resized crop is sampled back (bicubic) once per picture pixel, so the edit in those cases is as soft as the canvas it was painted on. Untouched pixels stay bit-identical there too, and in the feathered band your full-size picture replaces the canvas's softer copy of it.

How the empty check works

The new area is cut into 16 px blocks, and only blocks fully inside it are judged. A block counts as unpainted when it is within 20/255 of the canvas fill and flat. That is wide enough to catch a gray fill the model hands back a little darker. Painted scenery has texture even when it is dark, so a night street is not flagged. The same test runs on your picture, and only what the new area has beyond that counts: a black night sky may be continued as black. Up to nine frames spread across a clip are checked. Crop For Inpaint crops are not checked: their masked pixels are the old picture, not a fill.

How blend in works

• Where the edge is. A Crop + Rotate + Pad stitcher carries its generated-area mask; pixels at 0.98 or more are new area, so a turned picture's empty corners count too. A Load Image + Pad stitcher carries the rectangle the source sits in. From that, every pixel gets its distance to the picture's edge, positive inside and negative outside. A side of the picture that meets the frame is not an edge. The map is worked out once per stitcher and shared by every frame, with scipy's distance transform when it imports (ComfyUI requires scipy) and an exact torch fallback, good to 128 px, when it does not.
• Two layers. Both pictures are split into colour and tone (a Gaussian blur, sigma 8 px) and fine detail (what the blur takes away). Your picture's colour layer only reads pixels at least 4 px inside the edge, so fill that a turn or a resize mixed into the outermost pixels never leaks in.
• Two hand-overs. Detail changes over from the model's picture to yours across 10 px, starting at three quarters of the depth where the model's mask fell to 0.1: about how deep the model was free to redraw, and a little before its version lines up with yours (at least 8 px in; 13 px is assumed when the mask has no feather to read). For a Crop + Rotate + Pad stitcher that mask is its generated-area mask; for Load Image + Pad it is the feathered padding mask. Colour and tone change over slowly, on a smooth curve from 4 px to 14 px past the end of the detail hand-over.
• Texture. A straight cross-fade of two unrelated textures is weaker halfway across. Blend in puts 60% of the lost strength back.
• Exactness. Deeper than both hand-overs the result is your picture, bit for bit. Within 4 px of the edge and beyond it, it is the model's picture, bit for bit. The Crop + Rotate + Pad nodes' Blend and Grow paste (stitch_blend, stitch_grow) shape only the classic paste.
• Cost. Frames go through in chunks of about 2 MP, with a cancel check and progress between chunks: about 0.2 to 0.5 s per 1.5 MP frame on a desktop CPU.

Tone match under blend in

• What is compared. Tone layers (a Gaussian blur, sigma 5 px) of your picture and of the model's, both read over the same picture pixels at least 4 px inside the edge, so the grey rim a turn or a resize leaves there, and the new area's own content, never enter the reading. The band is the picture pixels where the sampler mask is above 0.02; highlights at 0.98 or more are left out. A clip is read through the mean of its frames.
• The drift model. Per RGB channel the difference is fitted as a curve over your picture's brightness, split into what a pinned pixel still drifts and what a free one does, weighted by the (equally blurred) sampler mask: d = A(v) + m (E(v) - A(v)). Both are piecewise linear on ten brightness points, denser in the shadows, with smoothness penalties, and fitted robustly (three reweighting passes). E is the drift in the new area. Klein re-renders a pinned picture, so its A is large; Krea keeps a pinned picture exact; Qwen redraws everything alike.
• The check. The band is cut into 96 px tiles in a checkerboard. The curves fitted on one colour must predict the other at least 15% better than no correction, both ways, or nothing changes. A local gain along the edge (sigma 48 px, at most 5%) is added only when it predicts another 5% better. A gain leaves black black, where an offset would lift it.
• Taking it off. Each value of the model's picture is taken back through its curve (v + drift(v) = value, three fixed-point steps), with the pixel's own mask mixing pinned and free, then divided by the local gain. The corrected picture goes through the blend in unchanged, so everything under "How blend in works" still holds.
• Cost. About 0.4 to 2 s per 1.5 MP stitch on a desktop CPU, once per stitch whatever the number of frames.

Measured

Krea 2 with AnyPaint on seven outpaints at two seeds each, 1.5 MP, feather 12: six pictures turned 5 to 15 degrees (one of them also padded on two sides) and one padded on three sides. "Colour step" is how much the stitch changes the model's picture just inside the edge compared with just outside it, so the scene itself does not enter it. Ranges are over the seven pictures, each the mean of its two seeds.

	classic, Tone match 1	blend in
colour step across the seam (median dE)	0.8 to 4.6	0.11 to 0.21
new area moved off what the model painted (dE)	0.9 to 3.0	0.00
fine detail lost in the strip at the seam	0 to 34%	0 to 22%
your picture exact from (px inside the edge)	20 to 56	23 to 38

The outer 20 to 35 px of your picture is the model's redraw either way, about the same share as classic. Blend in changes where that strip meets your picture, not how wide it is.

Tone match under blend in, on nine outpaints whose real continuation is known (Klein 9B and Krea 2, straight pads, feather 16 to 64): how far the new area's tone next to the edge is from the real photo, median dE.

	classic, Tone match 1	blend in, Tone match 0	blend in, Tone match 1
mean over the nine	1.98	2.57	1.80
Klein, feather 64 (three)	1.78 to 2.06	1.42 to 2.40	0.73 to 1.84
Krea 2, feather 32 to 64 (four)	1.76 to 2.14	2.84 to 3.19	1.41 to 2.33

On 18 real runs (Qwen Image 2.1, Klein 9B and Krea 2; night, daylight and flat backdrops; turned 5 to 15 degrees, straight pads on one, two or four sides, 4:5 and 16:9 ratio chips) no correction showed a seam line. Qwen drifted little: 4 of 7 runs got no correction and 3 got 0.4 to 0.6 dE. Klein got 1.8 to 6.3 dE, taking back a darkening or brightening of its own, and Krea 2 got 0.6 to 1.4 dE. Classic Tone match moved every one of those new areas by 1.5 to 6.9 dE; on a turned night scene it lifted the new area 2.8 L off the grey rim, where Qwen's real drift was 0.2.

Wiring

Crop For Inpaint ── image/mask ──> LaMa Inpaint ── image ──> Stitch Inpaint ── image
        └── stitcher ────────────────────────────────────────────┘