TTP_Image_Assy

This is the back half of TTPlanetPig's original tiled-upscale workflow - the pack's own README calls it "TTP Tile 1.0" now that Smart Tile 2.0 exists, but it's still the straightforward, reliable way to do a classic fixed-grid tiled upscale on Flux, Hunyuan, SD3, or anything else. If you cut an image into tiles with TTP_Image_Tile_Batch, ran each one through an upscaler or img2img sampler, this is the node that puts the pieces back together without visible seams.

How it works. The README describes it plainly: it "reassembles image tiles back into a complete image while preventing visible lines between the tiles," operating in pixel space (not latent space - you're compositing finished images, not blending noisy latents). It needs the same four pieces of information TTP_Image_Tile_Batch produced: tiles (your processed image batch - swap in individually-edited tiles here if you need to), positions, original_size, and grid_size, all "paired with" the tile batch node, plus a padding value (default 64) that controls how much the pieces overlap and get blended where they meet.

The one setting you'll actually touch is padding. It's the seam-hiding budget: too low and you'll see hard edges where tiles meet, especially if your upscaler or sampler introduced any per-tile color or texture drift; too high and you're throwing away real detail near the boundary. 64px is a sane starting point for a 1024px tile; scale it up if your tiles are much bigger.

Why this workflow instead of a fancier one. ControlNet Tile plus tiled diffusion is still, per the community, one of the most reliable photorealistic upscaling recipes around - it's not flashy, but it's predictable and it works on limited VRAM because you never hold the full-resolution image in the sampler at once. This pair of nodes (TTP_Image_Tile_Batch → sampler → TTP_Image_Assy) is that recipe with the plumbing already built for you.

Install. ComfyUI Manager, search "Comfyui_TTP_Toolset", or clone it manually: cd ComfyUI/custom_nodes && git clone https://github.com/TTPlanetPig/Comfyui_TTP_Toolset, then restart. Nothing extra to download for this node itself - no models, no heavyweight dependencies. If you want the full classic recipe, you'll separately need a ControlNet Tile checkpoint that matches your base model family (SD1.5, SDXL, or a newer architecture's tile model), which is a model download, not a pack dependency.

Common issues. The most common failure mode with any tiled reassembly, TTP's included, is visible seams - and it's almost always upstream of this node, not in it. If your tiles were sampled without something anchoring them to the source image (a Tile ControlNet at a sane strength, or a low enough denoise on img2img), each tile can hallucinate slightly different detail near its edges, and no amount of padding blending fully hides that divergence. The community's own summary of this exact mistake: skipping ControlNet Tile on a tiled upscale lets tiles diverge from the source and creates patchwork artifacts. If seams persist after raising padding, that's the signal to go back and fix the sampling step, not to keep cranking this node's blend width.

Also double-check that positions, original_size, and grid_size actually came from the same TTP_Image_Tile_Batch run that produced your tiles - if you regenerate the tile batch (say, with different tile_width/tile_height) but forget to reconnect all three, the reassembly math won't match and you'll get a garbled or misaligned output.