Artoriuz/ArtCNN

What it solves

It provides a collection of Single Image Super-Resolution (SISR) models specifically designed to upscale and clean anime content, addressing common issues like JPEG artifacts and chroma reconstruction reconstruction.

How it works

The project offers two distinct neural network architectures:

• C-series (Speed): Lightweight convolution layers with a single long-skip connection, optimized for real-time tasks like video playback. These are available as GLSL shaders and ONNX models.
• R-series (Quality): Deeper architectures using residual blocks and short-skip connections with more filters, designed for high-quality non-real-time rescaling.

Who it’s for

Users looking to upscale anime visuals, specifically those using video players like mpv or video processing frameworks like VapourSynth.

Highlights

• Multiple Configurations: Offers models ranging from lightweight (C4F16) to high-quality (R16F96).
• Specialized Variants: Includes specific models for denoising, sharpening, softening, and cleaning JPEG artifacts (4:4:4 and 4:2:0).
• Chroma Reconstruction: Dedicated models for reconstructing chroma in 4:2:0 BT.709 YCbCr content.
• Broad Compatibility: Supports ONNX format and GLSL shaders for integration into playback software.