VideoGigaGAN<T>

Models & Types · AiDotNet.Video.Enhancement

VideoGigaGAN: towards detail-rich video super-resolution with large-scale GAN.

Inheritance: NeuralNetworkBase<T> → VideoNeuralNetworkBase<T> → VideoSuperResolutionBase<T> → VideoGigaGAN<T>

For Beginners

VideoGigaGAN is like a very talented speed-painter. While diffusion models gradually “develop” a high-res image from noise (slow but versatile), VideoGigaGAN directly “paints” detail in one stroke – much faster. A special “detail shuttle” ensures real details are preserved and fake ones aren’t invented, while anti-aliasing prevents the annoying flickering between frames.

Usage:

How It Works

VideoGigaGAN (Xu et al., CVPR 2025) is the first large-scale GAN for video SR:

  • GigaGAN backbone: upscaled StyleGAN architecture with 1B+ parameters, generating

exceptional spatial detail in a single forward pass (faster than diffusion models)

  • Feature propagation with anti-aliasing: temporal feature propagation uses anti-aliased

flow warping to prevent temporal aliasing artifacts that cause flickering

  • High-frequency shuttle: a dedicated parallel pathway that extracts and preserves

genuine high-frequency details (edges, textures, text) from the input, preventing the generator from hallucinating false details while maintaining real ones

  • Temporal discriminator: a 3D discriminator evaluates both per-frame quality and

temporal consistency, penalizing flickering and motion artifacts

  • Supports up to 8x upscaling with rich perceptual details

Native mode implements the paper’s generator: a style-conditioned spatial backbone with temporal inflation, anti-aliased flow-guided recurrent propagation, and the high-frequency shuttle. ONNX mode remains available for importing published generator checkpoints.

Reference: “VideoGigaGAN: Towards Detail-rich Video Super-Resolution” (Xu et al., CVPR 2025)

Constructors

Constructor Summary
VideoGigaGAN(NeuralNetworkArchitecture<T> architecture, VideoGigaGANOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates a VideoGigaGAN model in native training mode.
VideoGigaGAN(NeuralNetworkArchitecture<T> architecture, string modelPath, VideoGigaGANOptions options) Creates a VideoGigaGAN model in native training mode.

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Upscale(Tensor<T> lowResFrames)