🏗️ Slice Combiner Hub

Slicing an image is the easy half. Putting the strips back together without a visible seam down the middle of every join is where tile-and-stitch workflows fall apart. Slice Combiner Hub is the "merge" end of Shima's sliced ecosystem: it takes up to ten individual slice images (or one batch), aligns them, and merges them with feathered alpha-blending into a single composite.

It's deliberately the mirror image of Sliced Image Hub. The Image Hub splits a batch into separate S1–S10 streams; the Combiner Hub accepts those same streams and glues them back. If you didn't fan anything out - you just have a batch from the Sliced Upscaler - wire image_batch in directly and it splits internally, then merges. Either path lands on the same merge logic.

The inputs

• orientation - Vertical or Horizontal; must match how the slices were produced.
• overlap - the same overlap value used when slicing. The merge relies on it: each subsequent slice is placed back at previous_position - overlap, and the overlap zone is where blending happens.
• feather_size and feathering - blend width and master toggle. Feathering off forces overlap to 0 (hard-edged stacking, the "sliced but not blended" mode).
• image_batch - one batch input alternative to the individual streams.
• S1–S10 - the ten individual slice ports.
• sliced_commons - the shared settings bundle; when connected, its orientation, overlap, feather_size, and feathering override the widgets. This is the wire that keeps your merge in sync with your upscaler.

Output: one image. It also reports the orientation, slice count, overlap, and final resolution into a used_values display on the node - handy when the merge result looks wrong and you want to see what it actually assembled.

How the merge works

The node center-pads any slices that have mismatched dimensions (a slice that came back smaller gets padded symmetrically so nothing shifts off-axis), then walks the slices in order. The first lands at position zero; each following slice is written into the canvas at position - overlap, and in the overlap zone a linear gradient blends the incoming strip over the existing canvas. That gradient is capped at min(feather_size, overlap), which is a smart guard - it stops the blend from extending past the region where the two slices actually share pixels.

Installing it

Part of KDB-USJP/shima_wf:

cd ComfyUI/custom_nodes
git clone https://github.com/KDB-USJP/shima_wf

Restart ComfyUI. Pure torch/pillow, no extra dependencies.

The honest advice

For one-shot upscales, the Sliced Upscaler's built-in merge (its image output is the merged result) is enough and you don't need this hub at all. The Combiner earns its place when you want to process slices independently - per-slice detailing, per-slice LoRA, per-slice anything - and then reassemble. If your merged image shows hard seams anyway, the usual culprit is an orientation mismatch or an upstream node that didn't pass the same overlap through - that's exactly what the sliced_commons wire is supposed to prevent.