DOGMA v35 Exact-Core Stitch

Every stitch node in this pack answers the same question - how much of the generated crop do I actually paste - and this one answers it with a split. The middle of the mask goes in at 100%, no matter what you set. The feathered edge is the only part the strength dial touches.

Why you'd want that

The knob people usually want is "less confident, please." You sampled a crop, it came back slightly different in tone, and you'd like to blend it in gently. But on a normal stitch, turning strength down to 0.3 means you also blend the middle of the mask at 30% - and the middle is the part you were trying to fix. The car with the fused wheels stays 70% fused.

Splitting the alpha solves it. Inside the core, the generated pixels win outright; in the feather, you get the ramp you asked for. The docstring states the other half of the guarantee too: no generated pixels appear outside the outward feather. Combined with the fact that the node never passes the base image through a VAE, that's the "unmasked pixels must be untouched" property that makes masked inpainting worth keeping in 2026 at all - an instruction-editing model re-encodes the whole frame, and unmasked drift compounds over a chain of edits.

The mechanism, exactly

For each crop entry:

1. Resample the patch to the recorded rectangle (bicubic, antialiased) and the mask to the same size (bilinear).
2. core = (mask >= 0.999), feather = mask - core.
3. alpha = clamp(core + feather * strength, 0, 1).
4. Composite: patch * alpha + region * (1 - alpha).

So the only thing strength does is scale the soft rim. 0.0 gives you a hard-edged paste of the solid core and nothing else - which, if your mask's feather came from a dilate-plus-blur mask node, is a perfectly sane setting. 1.0 is a straight masked composite.

There is no colour matching and no global correction, deliberately. Colour drift is a symptom of a VAE round-trip or a mismatched checkpoint, and patching the symptom by shifting the tile's mean moves the error rather than removing it.

Inputs and outputs

• base_image (IMAGE) - your master. The node clones base_image[0], so it's one image, not a batch.
• patches (IMAGE, list) - the sampled crops, in stitch order.
• masks (MASK, list) - one per patch, same order.
• stitch (DOGMA_STITCH, list) - the geometry entries a crop node produced. DOGMAClusteredMaskCrops emits exactly this.
• strength (FLOAT, default 1.0, 0–1, step 0.01) - feather weight only.

One output: image (IMAGE). Wire it into a preview or an image-list-to-batch node if you've been working in lists.

The node is INPUT_IS_LIST=True and iterates min() of the three list lengths, so mismatched lists truncate in silence. Count first, debug later.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/axior/ComfyUI-DOGMA-Nodes

Or ComfyUI Manager, searching DOGMA Nodes - it's in the Comfy registry as comfyui-dogma-nodes, so comfy node install comfyui-dogma-nodes also works. The pack has no Python dependencies: requirements.txt is a comment line, pyproject.toml declares dependencies = [], and the imports are torch, numpy and comfy.utils. Nothing to download, MIT licensed, and it doesn't care which model produced your patches.

Where people get burned

The >= 0.999 threshold is stricter than it looks. A mask that came out of a bilinear resample, a soft SAM head, or any blur pass may never actually reach 1.0 anywhere - in which case your "core" is empty, everything is feather, and strength behaves like a normal global opacity dial. Same symptom, different cause, and the fix is upstream: run the mask through a hard threshold (or a mask-builder node that produces an opaque core) before the stitch. DOGMAFastDualMaskV34 exists for this and its docstring says so - "opaque core feathered stitch mask."

Second: this is v35 in a pack that keeps superseding itself. The v10 sibling applies strength uniformly; v39's DOGMADetailBandStitchV39 takes the same four inputs plus four strength/drift dials and merges frequency bands instead of doing a straight paste. If you want a fast, predictable, one-number stitch, this is the one. If you want the generated tile's texture without letting it move your low-frequency geometry, that's the v39 one.