Tile Compositor (Merge)

The name tells you exactly what it does: this is the node that puts things back. Crop-and-refine is the most reliable trick in the ComfyUI playbook - grab a region, run a fresh sampling pass on it, paste the result where it came from - and Tile Compositor (Merge) is the third step. You'll recognize the pattern from Impact Pack's FaceDetailer, which made "detect, crop, resample, paste back" the standard practice. The difference here: instead of a detector finding the region, you pick it, and this node remembers where it was.

What makes it work is the pack's shared TILE_DATA contract. A crop source (usually this pack's Visual Marquee) hands you a crop and a little blob of metadata recording where that crop came from and what size it was. Tile Compositor reads that metadata and does the paste-back, so you never have to touch coordinates by hand.

How it works

Under the hood it's a masked blend, not a naive pixel copy. The node reads original_bbox (x, y, width, height) out of the tile_data, then:

• Auto-resizes the tile. If your processed tile doesn't match the crop dimensions (it usually won't after a latent/img2img pass), it Lanczos-upscales or downscales it to fit. No manual dimension matching.
• Builds a soft mask. It creates a mask over the region and blurs its edges with an average pool whose radius is your feather value, so the tile blends into the background instead of having a hard rectangular border.
• Composites. Merges the tile over the original image using that feathered mask, clipping everything to image bounds so you can't paste off-canvas.

One behavior worth knowing: if the tile_data you feed it is missing or invalid, it silently returns the base image unchanged. No error, no fanfare - the graph just does nothing. That's the safety valve, but it's also how a wrong wiring sneaks past you.

The inputs that matter

Only four, and three of them are obvious:

• base_image - the original image the tile came from.
• processed_tile - the result of whatever you did to the crop (sampler output, img2img pass, upscale).
• tile_data - the metadata from the crop source. Wire it straight from Visual Marquee (or anything else that emits the pack's TILE_DATA).
• feather (default 16, range 0–100) - the only knob you'll actually tune. 0 gives a hard seam, which you almost never want. 16 is a sensible starting blur; crank toward 100 for a very soft fade, and keep in mind very high values will visibly ghost the edge over detail in the background.

The single output, composite_image, is the original image with the processed tile blended in - exactly what you'd send to a Save Image node.

Getting it installed

It ships in the ComfyUI-Flow-Assistor pack, so install is pack-wide:

cd ComfyUI/custom_nodes
git clone https://github.com/Merserk/ComfyUI-Flow-Assistor.git

Then restart ComfyUI. Or search "Flow-Assistor" in ComfyUI Manager and hit install. The pack is V3-only (needs a current ComfyUI build with the comfy_api.latest API and Python 3.10+), has zero extra pip dependencies, and needs no model downloads for this node.

Common gotchas

The feather is where people get burned: at 0 the seam is ugly, but slamming it to 100 on a high-res composite gives you a soft, smeared halo. Start at the default and move it in small steps. The other trap is feeding it a stale or foreign tile_data - remember the silent-passthrough behavior, so check that your crop source is actually connected if your "edited" image comes back looking untouched.

It pairs with Visual Marquee (pick region → process → paste back), and if you're just sketching a graph without an interactive step in the way, the pack's Tile Manager can stand in as a dummy crop source that keeps the wiring valid.