Upscale by committee: one diffusion tile at a time

What it does

A ComfyUI node pack built around one idea: slice a large image into a grid, extract a text prompt for every tile via an image interrogator, upscale each patch with DIT models like Flux or Hunyuan, and stitch the results back together. It is essentially a memory-management strategy disguised as a workflow.

The interesting bit

Rather than blasting one global prompt at a massive canvas, the nodes feed localized conditions back into a merge stage that reassembles the image with configurable padding to hide seams. That per-tile granularity is what the authors claim keeps hallucinations down when you push far beyond native resolution.

Key highlights

• Targets DIT-based models: Flux, Hunyuan, and SD3.
• Assembly node runs in pixel mode and blends tiles with padding to suppress visible lines.
• Includes a TeaCache sampler node (ported from external repos) that roughly doubles Hunyuan Video generation speed on a 4090, at the cost of some fidelity and motion dynamics.
• Supports both pixel-space and latent-space workflows, plus ControlNet Tile integration.

Caveats

• The TeaCache speedup is explicitly flagged as a quality trade-off; faster renders can mean flatter dynamics.
• Several nodes (Cond to Batch, Coordinate Splitter) appear to exist mainly to wire internal plumbing together, so the setup is not a single drop-in node.
• The README is enthusiastic but thin on technical depth; exact memory savings or maximum tile counts are not stated.

Verdict

Worth a look if you are already deep in ComfyUI and need to push DIT upscaling past the VRAM wall. If you are not running tiled workflows or prefer end-to-end tools, this will feel like a bag of adapters.