This repository provides tiled samplers for ComfyUI, enabling users to generate larger images than typically supported by memory constraints. It targets users of ComfyUI, particularly those working with high-resolution image generation, by offering a solution to overcome VRAM limitations.

How It Works

The core innovation lies in its tiling strategies, which break down large images into smaller, manageable tiles for denoising. The "random" strategy minimizes seams by denoising tiles step-by-step with randomized positions, effectively blending them over time. A "random strict" variant uses masking to prevent border tile cropping. The "padded" strategy uses padding to provide context to each tile, reducing seams at the cost of increased tile processing.

Quick Start & Requirements

• Install via ComfyUI custom nodes manager or by cloning the repository into the custom_nodes directory.
• Requires ComfyUI.
• See ComfyUI_TiledKSampler examples for visual demonstrations.

Highlighted Details

• Offers multiple tiling strategies: "random," "random strict," and "padded," each with different trade-offs for seam reduction and compatibility.
• "Random" strategy is noted for exceptional seam hiding, even from pure noise.
• "Padded" strategy is compatible with uni samplers but processes up to 4x more tiles.
• Roadmap includes support for latent masks, image-wide ControlNets/T2I adaptors, and area conditioning.

Maintenance & Community

• The repository is maintained by BlenderNeko.
• Roadmap items are tracked within the README.

Licensing & Compatibility

• The repository does not explicitly state a license in the provided README.

Limitations & Caveats

• The "random" tiling strategy is incompatible with uni samplers.
• The "random strict" strategy may not work well with SDE samplers.
• The "padded" strategy can require processing up to four times more tiles.
• Support for tile-wide ControlNets and T2I adaptors is still under development.