IQVFI<T>

Models & Types · AiDotNet.Video.FrameInterpolation

IQ-VFI: image quality-aware video frame interpolation with degradation adaptation.

Inheritance: NeuralNetworkBase<T> → VideoNeuralNetworkBase<T> → FrameInterpolationBase<T> → IQVFI<T>

For Beginners

Most frame interpolation methods assume input frames are clean and high-quality. IQ-VFI first checks how good each part of the input frames is, then adjusts its interpolation strategy accordingly. This means it works better on real-world videos that may have noise, blur, or compression artifacts.

Usage:

How It Works

IQ-VFI (2024) adapts interpolation based on input image quality:

  • Quality assessment module: estimates per-pixel quality scores for input frames using a

learned no-reference image quality assessment (NR-IQA) branch, identifying regions with noise, blur, compression artifacts, or other degradations

  • Degradation-adaptive flow: the optical flow estimation network receives quality maps as

additional conditioning, so it can be more conservative in degraded regions (where flow estimation is unreliable) and more aggressive in clean regions

  • Quality-guided fusion: the blending weights between warped frames incorporate quality

scores, favoring the higher-quality frame contribution in each spatial region

  • Quality-aware training: training uses a quality-stratified sampling strategy that ensures

the model sees diverse degradation levels and learns robust interpolation for each

Reference: “IQ-VFI: Image Quality-Aware Video Frame Interpolation” (2024)

Constructors

Constructor Summary
IQVFI(NeuralNetworkArchitecture<T> architecture, IQVFIOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates an IQ-VFI model in native training mode.
IQVFI(NeuralNetworkArchitecture<T> architecture, string modelPath, IQVFIOptions options) Creates an IQ-VFI model in native training mode.

Properties

Property Summary
IQVFIDistillation<T> Distillation { get; } Gets the knowledge-distillation objective: pyramid reconstruction plus acceleration and motion distillation from a privileged teacher, gated by the selective mask.
ImplicitQuadraticMotion<T> QuadraticMotion { get; } Gets the implicit quadratic motion model, which modulates linear intermediate flows into quadratic ones using a latent acceleration prior.

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Interpolate(Tensor<T> frame0, Tensor<T> frame1, double t)