CALLIC: Content Adaptive Learning for Lossless Image Compression

TL;DR

This work tackles the amortization gap in learned lossless image compression by uniting the Minimum Description Length principle with Parameter-Efficient Transfer Learning. It introduces CALLIC, a content-adaptive framework built on the Masked Gated ConvFormer (MGCF), Cache then Crop Inference (CCI), and Rate-guided Progressive Fine-Tuning (RPFT) to adapt models per image with minimal overhead. The approach achieves state-of-the-art compression performance across diverse datasets while maintaining lightweight models (MGCF ~575K parameters; CALLIC adds ~25K mergeable weights) and offers controllable trade-offs between encoding speed and accuracy. By jointly optimizing the incremental weights and the image stream under MDL, CALLIC delivers practical, per-image adaptive lossless compression that generalizes well to out-of-domain content.

Abstract

Learned lossless image compression has achieved significant advancements in recent years. However, existing methods often rely on training amortized generative models on massive datasets, resulting in sub-optimal probability distribution estimation for specific testing images during encoding process. To address this challenge, we explore the connection between the Minimum Description Length (MDL) principle and Parameter-Efficient Transfer Learning (PETL), leading to the development of a novel content-adaptive approach for learned lossless image compression, dubbed CALLIC. Specifically, we first propose a content-aware autoregressive self-attention mechanism by leveraging convolutional gating operations, termed Masked Gated ConvFormer (MGCF), and pretrain MGCF on training dataset. Cache then Crop Inference (CCI) is proposed to accelerate the coding process. During encoding, we decompose pre-trained layers, including depth-wise convolutions, using low-rank matrices and then adapt the incremental weights on testing image by Rate-guided Progressive Fine-Tuning (RPFT). RPFT fine-tunes with gradually increasing patches that are sorted in descending order by estimated entropy, optimizing learning process and reducing adaptation time. Extensive experiments across diverse datasets demonstrate that CALLIC sets a new state-of-the-art (SOTA) for learned lossless image compression.

Figures (4)

• Figure 1: Overview of the proposed architectures and mechanisms in CALLIC: We first simplify the attention mechanism using a content-adaptive convolutional gating mechanism (a). Building on this, we introduce the Masked Gated ConvFormer (MGCF) (b). To accelerate the coding process, we propose Cache then Crop Inference (CCI) (c), which caches activations before masked convolution layers and then performs convolution on cropped features at coding positions. For instance-level content adaptivity, we fine-tune the pre-trained model on the test image. This involves adding learnable parameters through low-rank decomposition of MGCF layers (Right), fine-tuning these additional parameters with the proposed Rate-guided Progressive Fine-tuning (RPFT), and encoding them into the bitstream along with the image (Left).
• Figure 2: Illustration of incremental curves, $T=50$. The shape of the curve is controlled by parameters $b, d, e$.
• Figure 3: Steps and performance trade-off for CALLIC. Faster encoding can be achieved with minimal performance trade-offs.
• Figure 4: Different configurations for RPFT on Kodak. We ablate different increasing curves and sorting orders in this experiment.