ComfyUI SeedVR2 Tiling Wrapper

A complementary tiling implementation for the official SeedVR2 node package by Numz. This set of nodes is designed to bring advanced, VRAM-aware image tiling, seamless stitching, perceptually accurate color matching, and artifact-free sharpening to your ComfyUI workflows.

Features & Nodes

1. SeedVR2 Tile Splitter (VRAM Aware)

Splits high-resolution images into manageable overlapping tiles to prevent Out-of-Memory (OOM) errors during heavy processing (like upscaling or img2img).

• Dynamic VRAM Scaling: Automatically detects available VRAM (8GB, 12GB, 16GB+) and adjusts the base tile size accordingly, or lets you override it manually.
• Luminance-based Noise Injection: Optionally injects noise into tiles to help diffusion models add detail, scaling the noise by luminance to avoid destroying pure black areas.

2. SeedVR2 Tile Stitcher (Seamless)

Reconstructs the image from processed tiles without visible seams or grid artifacts.

• Laplacian Pyramid Blending: Instead of simple linear alpha blending, this node uses multiresolution pyramid blending. It breaks the tiles into different frequency bands (Laplacian pyramid) and blends them using a Gaussian pyramid mask, ensuring both low-frequency gradients and high-frequency details merge flawlessly.
• Black Level Fix: Includes a parameter to optionally shift and correct raised black levels that often occur after complex diffusion processing.

3. Advanced Color Match (OKLAB MKL)

A state-of-the-art color matching node designed to transfer the exact color palette from a source image to a target image.

• OKLAB Color Space: Operates entirely in the perceptually uniform OKLAB color space, which mimics how the human eye perceives color, preventing the hue shifts common in RGB/HSV matching.
• Monge-Kantorovich Linear (MKL) Optimal Transport: Uses statistical Optimal Transport theory to match the covariance matrices of both images. This ensures a mathematical and perceptually accurate transfer of color distributions.
• Luma Preservation Mode: Option to only match the 'a' and 'b' chrominance channels, preserving the original lighting and structural depth (L channel) of the target image.

4. CAS Luma Sharpening

An advanced sharpening technique that enhances details without artificial halos or color bleeding.

• YCbCr Luma Isolation: Converts the image from RGB to YCbCr and exclusively targets the Luma (Y) channel. This ensures that sharpening only affects structure, leaving color channels completely untouched.
• Contrast Adaptive Sharpening (CAS): Adapts the sharpening strength based on local contrast to prevent over-sharpening already sharp edges, originally adapted from AMD's FidelityFX.

Installation

1. Navigate to your ComfyUI custom_nodes directory.
2. Clone this repository:

   git clone https://github.com/shikasensei-dev/ComfyUI-SeedVR2-TilingWrapper.git
   

3. Restart ComfyUI. The nodes will appear under the SeedVR2_Tiling category.

Dependencies

No external dependencies outside of a standard ComfyUI environment are required, as it relies purely on native torch and torch.nn.functional operations.

Technologies Used

• PyTorch (for fast, GPU-accelerated tensor operations, border padding, and SVD decomposition)
• Monge-Kantorovich Linear (MKL) Optimal Transport
• Laplacian & Gaussian Image Pyramids
• OKLAB & YCbCr Perceptual Color Spaces
• AMD FidelityFX Contrast Adaptive Sharpening (CAS) algorithm