From c4b5251a9d0d83f031ba17aca131119a5b8e01d9 Mon Sep 17 00:00:00 2001 From: George Pearse Date: Fri, 21 Nov 2025 17:07:54 +0000 Subject: [PATCH] Add SOTA compression techniques overview to README Document current state-of-the-art approaches including ELIC, Cheng2020 with checkerboard context, variable bitrate models, and the evolution of context modeling techniques. --- Readme.md | 23 +++++++++++++++++++++++ 1 file changed, 23 insertions(+) diff --git a/Readme.md b/Readme.md index 5ddf644..b686a2a 100644 --- a/Readme.md +++ b/Readme.md @@ -18,6 +18,29 @@ CompressAI currently provides: ![PSNR performances plot on Kodak](assets/kodak-psnr.png) +## SOTA Techniques + +### Image Compression + +1. **ELIC (2022)** - State-of-the-art with: + - Uneven channel groups [16, 16, 32, 64, 192] + - Space-channel context (SCCTX) + - Non-local attention + - ResidualBottleneckBlock (1x1→3x3→1x1) +2. **Cheng2020 with Checkerboard Context** - Combines GMM attention model with checkerboard 2-pass decoding for parallelism +3. **JointAutoregressiveHierarchicalPriors (mbt2018)** - Best compression via combined hyperprior + autoregressive context + +### Variable Bitrate + +- **VBR models (2024)** with learnable gain parameters and quantization-reconstruction offsets + +### Video Compression + +- **ScaleSpaceFlow (ssf2020)** - Multi-scale Gaussian pyramid with optical flow + scale field prediction + +### Context Modeling Evolution (newest = most advanced) + +1. Factorized prior → 2. Hyperprior → 3. Autoregressive → 4. Checkerboard → 5. Channel groups → 6. SCCTX (ELIC) > **Note**: Multi-GPU support is now experimental.