← Back to all projects

ATLAS

ATLAS project screenshot

A ground-up next-generation hybrid video codec combining classical signal processing with AI-assisted analysis, semantic modeling, and advanced multi-pass encoding. Features modular architecture with background reuse, semantic-aware compression, deterministic neural enhancement framework, and real-time GPU decoding. Includes full VLC/FFmpeg integration and automated VMAF testing dashboard.

Technology

backendc++pythoncmake
databasesqlite3sqlalchemy
devopsdockerci/cd
toolsffmpegvlcvmafmetalcudavulkan

🌐 ATLAS — Next-Generation Hybrid Video Codec

Long-Form Project Summary for KNWS Showcase

ATLAS is a ground-up video codec designed to explore the boundaries of modern compression by combining classical signal-processing techniques with AI-assisted analysis, semantic modelling, and advanced multi-pass encoding. The project spans a full production toolchain: custom codec libraries, VLC and FFmpeg integration, GPU-optimised components, automated benchmarking, and a complete test infrastructure.

ATLAS was built to answer a single question: How much can a codec improve when it understands the structure, meaning, and behaviour of the video itself?

🚀 Project Goals

  1. Design a new hybrid codec architecture that supports classical DCT + motion vectors while layering modern techniques such as semantic analysis, background reuse, and deterministic neural enhancement.
  2. Achieve competitive or better performance compared to H.264/H.265 across a wide range of content.
  3. Provide real-world tooling so the codec can be used, tested, and validated inside standard video processing ecosystems.
  4. Establish a development platform capable of supporting experimental compression layers without breaking determinism.

🧱 Core Architecture & Design

ATLAS consists of multiple modular layers, each of which can be enabled, disabled, or combined depending on the encoding scenario. The architecture was intentionally designed so new layers can be added without destabilising the core codec engine.

1. Classic Codec Backbone

A full traditional encoder/decoder pipeline, including:

  • I/P/B frames
  • Block-based transforms
  • DCT and quantisation
  • Motion search & compensation
  • Temporal prediction
  • rANS entropy coding

This provides a deterministic, battle-tested foundation that other layers build upon.

2. ATLAS Background Reuse Layer

A central innovation of the project.

ATLAS identifies static background regions, extracts them into an atlas texture, and reuses those blocks across frames with negligible additional cost. Only changes, deltas, or new tiles need to be encoded.

This dramatically reduces bitrate on scenes with stationary backgrounds:

  • Interview content
  • News broadcasts
  • Films with long static shots
  • Gaming footage
  • Any content with static geometry

The atlas is referenced by tiles, tracked over time, and compressed independently.

3. Semantic-Aware Compression (SEM)

A lightweight, codec-embedded semantic analysis system designed to identify:

  • Static regions
  • Moving objects
  • Shot boundaries
  • Scene changes
  • Object reappearance
  • Meaningful motion vs unimportant noise

These signals inform the encoder so it can:

  • reduce bitrate in visually irrelevant areas
  • allocate more bits to complex or moving regions
  • trigger intra refreshes at shot boundaries
  • track important objects more efficiently

SEM is a non-neural, deterministic feature extractor designed specifically for codec usage.

4. GEN — Deterministic Neural Enhancement Layer (Framework Complete)

GEN establishes a framework for future neural enhancement models such as:

  • detail recovery
  • texture hallucination
  • denoising
  • super-resolution

The key innovation is that GEN is deterministic, meaning:

  • the same input always produces the same reconstructed frame
  • the codec remains bit-exact
  • distribution across platforms is safe

(Models themselves are still pending; the framework, hooks, and build system are complete.)

5. TOK — Token-Space Residuals (Planned Layer)

This introduces vector-quantised residuals for:

  • ultra-low-bitrate scenarios
  • texture synthesis
  • perceptually-guided reconstruction

TOK provides a foundation for extremely compact residual encoding, especially useful for AI-assisted generation of fine textures.

6. TND — "Tomorrow Never Dies" Reverse-Pass Encoding

A novel multi-pass encoding strategy.

TND uses a reverse pass to analyse frames from the future so the encoder can:

  • pre-allocate bits more intelligently
  • prioritise upcoming complex motion
  • insert keyframes with future context
  • stabilise VBR fluctuations

This mode is ideal for offline encoding workflows.

7. Licensing & Watermarking Layer

Includes:

  • ATLS atom for embedding licensing metadata
  • Deterministic watermarking for forensic analysis
  • Ed25519 signing pipeline
  • Token-based identification

This allows each encoded file to carry tamper-proof metadata.

8. Real-Time GPU Decoding

Decoder-side optimisations for:

  • Metal (macOS)
  • CUDA (NVIDIA)
  • Vulkan (Linux/Windows cross-platform)

This ensures that ATLAS can run at high frame rates even at 4K+ resolutions.

🛠 Tooling & Integrations

ATLAS includes full integration into industry-standard video tools.

VLC v4 (Custom Build)

A custom distribution of VLC 4 with:

  • native ATLAS decoder module
  • tile atlas visualisation tools
  • debug overlays
  • extended media info for ATLS atoms

This allows real-world playback of ATLAS files.

FFmpeg (Custom Build)

FFmpeg was patched to add:

  • libatlas encoder
  • libatlas decoder
  • atlas-specific AVCodecContext fields
  • custom AVFrame metadata
  • seamless integration with filters and tools

This enables:

  • command-line encoding
  • piping from other tools
  • automated CI testing
  • benchmark comparisons with H.264/H.265

🧪 Automated Testing & Flask Web System

ATLAS includes a full Flask-based testing suite:

  • Upload or reference test videos

  • Automatically encode using ATLAS and H.264/H.265 baselines

  • Automatically run Netflix VMAF

  • Produce charts, deltas, data tables, regression comparisons

  • Track improvements or regressions across versions

  • Run custom scenarios such as:

    • static backgrounds
    • high-motion content
    • animation
    • colour-sensitive scenes
    • artificial stress tests

The dashboard is used daily for development and performance tuning.

🧰 Build System & Cross-Platform Support

ATLAS supports:

  • macOS (Intel + ARM)
  • Linux (x86 + ARM)
  • Windows (MSVC + MinGW)
  • GPU-enabled builds
  • headless encoder mode
  • CI-driven binary releases

The build uses a customised CMake + scripting pipeline that orchestrates:

  • Codec library builds
  • Plugin assembly for VLC and FFmpeg
  • GPU kernel builds
  • Static and shared-library variants
  • Release packaging
  • Regression test snapshots

📈 Outcome & Impact

The project demonstrates:

  • the ability to design a full-scale video codec
  • understanding of compression theory, GPU systems, and multimedia toolchains
  • end-to-end engineering from core algorithm design to distribution tooling
  • practical interoperability with industry tools
  • real performance testing against standard frameworks (Netflix VMAF)
  • a modular foundation capable of future research and extensions

ATLAS is already able to encode and decode real-world content, integrate with FFmpeg and VLC, run across multiple platforms, and serve as a research platform for advanced hybrid compression techniques.