PJ 图像切片原位还原器

So you sliced a big image into a grid, upscaled or re-inpainted one tile in glorious high resolution, and now you face the part everyone dreads: putting it back without a visible rectangle where you pasted. PJ_Image_Slice_Reassembler (pongjoo/ComfyUI-PJLatent, menu "PJ_Nodes/Image") is the return half of the pack's slice-edit-stitch loop - it reads the exact tile coordinates from the slicer, drops your edited tiles into their original spots, and feathers the seams with a cosine blend so the patch reads as part of the image rather than a sticker.

This is the same philosophy as the crop-and-stitch pattern that's become the community default for region edits: never pass unchanged pixels through a VAE, composite only what you actually changed. Here it's grid-shaped and manual - you decide which tiles deserve a redo.

Feeding it is two wires and a choice. Connect 基底原图 (the base image - the slicer's 原图 output or your original upstream) and 切片数据 (the slicer's coordinate JSON, verbatim; the input's tooltip is explicit about which port to use). Then say which tiles you changed:

• Direct slots: 块1_图片 … 块30_图片, mapped by index - wire your edited 块5_图片 result into 块5_图片 and it replaces tile 5. No per-slot target numbering to set; the port number is the tile number.
• 批次替换图片 + 批次对应块序号 - for many tiles at once. The order field is a small parser in its own right: "5", "2, 5, 8", "1-3, 5" (ranges and lists), or "全部". Images land on the listed tiles in sequence.

The two settings that decide quality:

• 边缘羽化像素 (feather) - the blend width in pixels; 0 is a hard edge. The tooltip's recommendation of 12–32 matches the community's mask-blur wisdom: too little shows seams, too much bleeds the patch into its neighbors. The feather is a cosine S-curve applied only on seams that touch another tile - edges sitting on the image's outer border stay crisp, so you don't get a faint halo around the whole reassembled image.
• 分辨率处理模式 (resolution mode) - three behaviors. Default scale the replacement back to the original slice size (keeps base resolution - right for a same-size inpaint). The standout is upscale the whole canvas proportionally to the highest tile magnification: if one tile came back 2× or 4× from a super-resolution pass, the entire composite grows to match and your upscaled detail survives instead of being shrunk away. Third is force-stretch the tile to fit regardless of aspect.

What you get back. Three outputs:

• 还原后图像 - the assembled image (IMAGE).
• 替换区域遮罩 - a MASK marking exactly the composited regions, which is genuinely useful if you want a follow-up pass to only touch where you just edited.
• 还原汇总信息 - a STRING report listing each replacement's coordinates, resolution handling, and which tiles were left untouched. Nice for audits of batch work.

It's also defensive where pasting gets ugly: RGBA replacements have their alpha channel multiplied into the blend (so a tile from a transparent/outpaint pipeline doesn't leave hard edges), and channel/RGB-RGBA mismatches are reconciled rather than crashing.

Honest caveats. No 切片数据 means an error string out the summary port - connect it before anything else. And it's index-based: if you slice with a different cut layout than the data describes, tiles land on the wrong coordinates, so keep the reassembler fed from the same slicer run (or its 原图/切片数据 pair) that produced the tiles.

Install

Part of pongjoo/ComfyUI-PJLatent (MIT). ComfyUI Manager → search "ComfyUI-PJLatent", or:

cd ComfyUI/custom_nodes/
git clone https://github.com/pongjoo/ComfyUI-PJLatent.git

Restart ComfyUI - pure torch/PIL, no models or extra deps. Look for PJ 图像切片原位还原器 under "PJ_Nodes/Image".