Reproduce “CALLIC: Content Adaptive Learning for Lossless Image Compression” on OpenResearch

About this paper

Daxin Li, Yuanchao Bai, Kai Wang, Junjun Jiang, Xianming Liu, Wen Gao

CALLIC introduces a content-adaptive learning framework for lossless image compression that fine-tunes a lightweight generative model to individual image characteristics. It achieves superior compression ratios across diverse datasets by consistently outperforming prior methods, while being significantly more parameter-efficient than other state-of-the-art learned codecs.

Key insights

• Content-adaptive learning, guided by the Minimum Description Length (MDL) principle and Parameter-Efficient Transfer Learning (PETL), effectively bridges the 'amortization gap' in learned compression by tailoring the model to each specific image.
• A lightweight architecture with efficient convolutional gating (MGCF) can achieve strong baseline performance and is crucial for making per-image model adaptation computationally viable.
• Optimized inference strategies like Cache then Crop Inference (CCI) and intelligent fine-tuning methods like Rate-guided Progressive Fine-Tuning (RPFT) are essential for balancing encoding speed and compression performance in content-adaptive codecs.

Summary by alphaXiv.