Generate Seam Mask

Tiled upscaling solves one problem and quietly creates another. When you want a 4K image out of a 1024px source on a 6GB card, you can't denoise the whole thing at once - you split it into overlapping tiles, run KSampler on each, and stitch them back together. The stitching is where it falls apart: every tile boundary can leave a visible "cut line" where the blend doesn't quite land. GenerateSeamMask exists to draw a mask of exactly those boundaries, so you can run a second, targeted denoise pass on the seams and only the seams. It's the patch that makes the patchwork invisible.

What it actually is

It's the single custom node shipped in ethanfel/ComfyUI_UltimateSGUpscale, a small pack that replicates the core of UltimateSDUpscale (the old tiled-upscaling workhorse, usually bundled in Impact Pack) as a transparent workflow built from ComfyUI's own built-in nodes. The whole pack is one node plus one example workflow - no model downloads, no API key, no requirements.txt. If you've seen the classic tiled upscale where the seams show, this is the seam-fix half of that recipe, rebuilt so you can inspect every step instead of trusting a black box.

How it works

The node reproduces SplitImageToTileList's tiling math: stride is tile_size - overlap, tiles walk across the image with that stride, and each adjacent pair of tiles shares an overlap region. The mask puts a band centered on the midpoint of each overlap - which is where the seam actually lives. In binary mode those bands are hard white and everything else is black. In gradient mode each band is a linear ramp that peaks at the seam center and fades to zero over seam_width pixels, which is what makes it usable with a DifferentialDiffusion node so seam centers get full denoising while the edges melt into the surrounding image.

The inputs that matter

Only a few, and only one real gotcha:

• image_width / image_height - just wire these straight from a GetImageSize node.
• tile_width / tile_height / overlap - these must match the tiling in your main pass. This is the one thing that burns people.
• seam_width (default 64) - how wide the fix band is. Too narrow and the seam pokes through; too wide and you're denoising parts that were fine.
• mode - binary for a hard mask fed through ImageToMask into SetLatentNoiseMask, gradient for the DifferentialDiffusion route.

The output is a single IMAGE: a mask with white bands at seam positions, black elsewhere.

Installing it

The pack is one node, so install is boring in the good way:

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI_UltimateSGUpscale.git

Restart ComfyUI and it shows up as Generate Seam Mask under image/upscaling. ComfyUI Manager can also find it by searching "UltimateSGUpscale". There are no model files to fetch. The one real dependency is a recent enough ComfyUI: it needs the built-in SplitImageToTileList and ImageMergeTileList nodes, added upstream in PR #12599. If you search your node list and those aren't there, update ComfyUI - don't blame the pack.

Troubleshooting

• The bands don't line up with the actual seams. You changed a tile size or overlap in pass one and didn't mirror it here. The node has no way to know what pass one did; it trusts you. Keep GetImageSize wired in and the two passes in lockstep.
• The mask is completely black. Your image fits in a single tile, so there are no seams - nothing to fix, and the second pass is wasted compute. Skip it.
• Hard edges even in gradient mode. The gradient only earns its name if a DifferentialDiffusion node is consuming it; fed to SetLatentNoiseMask on its own it still denoises a hard band.

The seam problem this fixes is as old as tiled upscaling - the classic A1111-era advice was "more overlap, more mask blur" to hide cut lines. This is the more surgical version: find the exact seam, fix only that. It's a niche helper, but if you're doing print-size upscales on a modest card, it's the difference between a clean result and one you keep zooming in on, frowning.