A Standard for High‑Quality, Low‑Complexity Neural Video Up‑sampling

The Up-sampling Filter for Video Applications (EVC‑UFV) V1.0 standard addresses a key challenge in modern video processing pipelines: restoring high‑quality video after resolution reduction. Typical video workflows include capture, down‑sampling, encoding, transmission, decoding, and up‑sampling before rendering. While traditional filters such as bicubic and Lanczos are widely used, they rely purely on mathematical interpolation and cannot recover lost details.

EVC‑UFV introduces a new paradigm: AI‑trained, non‑linear up‑sampling filters based on Super Resolution, capable of reconstructing higher‑quality video beyond conventional interpolation limits, while also providing tools to control computational complexity.

The EVC‑UFV Approach

The standard specifies how to design and deploy neural network–based up‑sampling filters trained on large video datasets. These filters learn to reconstruct high‑resolution content with significantly improved fidelity.

Two key capabilities define EVC‑UFV:

  • AI‑based up‑sampling design – Neural networks trained using Super Resolution techniques
  • Complexity reduction – Standardised pruning methods to reduce computational cost with minimal performance loss

Reference Model

EVC‑UFV defines a structured workflow enabling users to design interoperable up‑sampling solutions:

  • Training Dataset Preparation Selection of representative video sequences and generation of training data
  • Model Development
    • Pre‑training phase
    • Fine‑tuning phase
    • Definition of the up‑sampling neural network architecture
  • Deployment Model Standardised procedures to apply the trained model to up‑sample images and video sequences
  • Complexity Control A pruning algorithm reduces network size and execution cost while preserving performance

The standard includes reference software for both training and pruning, enabling reproducible and interoperable implementations.

Standardised Filters

EVC‑UFV V1.0 provides ready‑to‑use parameters for two practical scenarios:

  • Standard Definition (SD) → High Definition (HD)
  • High Definition (HD) → Ultra High Definition (UHD)

The HD→UHD filter can also be used for SD→HD conversion with limited performance loss, offering flexibility in deployment.

Performance Gains

EVC‑UFV delivers significant improvements over traditional interpolation:

  • Up to 22.5% BD‑rate reduction in VVC configurations
  • Consistent gains across HEVC and VVC, Low Delay and Random Access modes
  • Strong improvements across all objective quality metrics

Even after pruning:

  • ~40% reduction in model parameters
  • <1% performance loss compared to unpruned models

This demonstrates an optimal balance between quality and efficiency.

Tools and Implementation

The MPAI Git software platform enables users to:

  • Upload SD or HD images
  • Select pruned or unpruned filters
  • Generate and download up‑sampled images

This facilitates rapid testing, benchmarking, and deployment of EVC‑UFV filters.

Data and Interoperability

EVC‑UFV promotes:

  • Standardised design procedures
  • Reproducible neural network models
  • Interoperable implementations across applications

By defining both methodology and reference outputs (weights, datasets, algorithms), the standard ensures consistent performance across systems.

Open and Flexible Design

EVC‑UFV follows key MPAI principles:

  • AI‑driven processing with clearly defined workflows
  • Separation between design, optimisation, and deployment
  • Support for continuous innovation without breaking compatibility

This enables developers to improve models while remaining compliant with the standard.

Benefits

EVC‑UFV V1.0 enables:

  • Researchers to advance AI‑based video reconstruction techniques
  • Codec developers to integrate high‑performance up‑sampling into compression pipelines
  • System integrators to balance quality and computational cost
  • Industry to adopt interoperable, standardised AI solutions
  • Users to experience higher‑quality video at lower bitrates

Conclusion

EVC‑UFV V1.0 transforms video up‑sampling from simple interpolation to intelligent reconstruction, delivering substantial efficiency gains and visual improvements. By combining AI‑based design with standardised complexity control, it establishes a powerful and practical foundation for next‑generation video systems.