This repository provides a comprehensive collection of self-trained single image super-resolution (SISR) models, primarily targeting AI enthusiasts and researchers looking for advanced upscaling solutions. The benefit is access to a wide variety of specialized models, many trained with custom datasets and degradation handling, offering potentially superior results compared to generic upscalers.

How It Works

The project focuses on training SISR models using various architectures and custom degradation pipelines, often inspired by or extending the Real-ESRGAN framework. The author emphasizes meticulous dataset curation and experimentation with "on-the-fly" (OTF) degradation settings, including noise, blur, and compression, to create models robust to real-world image imperfections. This approach allows for specialized models tailored to specific image types (photos, anime, AI-generated art) and degradation patterns.

Quick Start & Requirements

• Models are primarily intended for use with the chaiNNer application, with a YouTube tutorial available for setup.
• Models are distributed as .safetensors files.
• A Python script is available for downloading models.
• Official releases are being organized via GitHub releases.

Highlighted Details

• Over 111 released models, with ongoing updates and GitHub releases.
• Extensive experimentation with OTF degradation parameters, detailed in the README.
• Includes custom datasets like BHI, NomosRealWeb, SSDIR, Lexica, FaceUp, and Nature.
• Performance benchmarks comparing various architectures are provided.

Maintenance & Community

• The project is maintained by a single author ("Phhofm").
• Links to the Enhance Everything! Discord Server are provided.
• Models are also being released on Hugging Face for easier integration.

Licensing & Compatibility

• The README does not explicitly state a license. Users should verify licensing for commercial or closed-source use.

Limitations & Caveats

• The project is primarily a personal collection and may lack formal support structures.
• Some older models are noted as potentially undertrained or having issues (e.g., detail loss with "R" models).
• The interactive comparison website is not up-to-date due to repository size.