TeamMoeAI/MoeSR

What it solves

MoeSR is a tool designed for image super-resolution (SR) and image restoration (IR) specifically tailored for ACGN (Anime, Comic, Game, and Novel) content, such as illustrations, CG, and manga. It aims to improve detail quality and clarity in a way that remains faithful to the original art style.

How it works

The software uses deep learning methods, including improved or self-developed models, to upscale and restore images. It supports various specialized models based on architectures like Real-ESRGAN, HAT (Hybrid Attention Transformer), and UFormer. The tool runs locally on Windows 10 (1903+) and utilizes DX12-compatible GPUs for acceleration, falling back to CPU processing if no compatible GPU is available.

Who it’s for

Artists, collectors, and enthusiasts of ACGN-style artwork who need to upscale low-resolution images or restore images that have been heavily compressed (e.g., from platforms like X or Pixiv).

Highlights

• Specialized ACGN Models: Includes support for various models tailored for Japanese illustrations and general 2D art.
• Image Restoration: Features the MoeIR model for recovering details from images with compression artifacts, scaling issues, and artifacts.
• Workflow Automation: Supports a workflow system with nodes for inference, scaling, and conditional logic for width/height.
• Broad Format Support: Handles multiple image formats including JPG, PNG, AVIF, and WebP.
• Asynchronous Processing: Implements asynchronous image saving to keep the GPU active and prevent idle time during file I/O.