CAT: Content-Adaptive Image Tokenization

TL;DR

The paper introduces Content-Adaptive Tokenizer (CAT), a dynamic image encoding method that adjusts token count based on image complexity, improving reconstruction quality and efficiency.

Contribution

It proposes a novel content-adaptive encoding scheme and a caption-based evaluation system to optimize image tokenization based on content complexity.

Findings

Improves FID score over fixed-ratio baselines.

Boosts inference throughput by 18.5%.

Demonstrates robust image reconstruction performance.

Abstract

Most existing image tokenizers encode images into a fixed number of tokens or patches, overlooking the inherent variability in image complexity. To address this, we introduce Content-Adaptive Tokenizer (CAT), which dynamically adjusts representation capacity based on the image content and encodes simpler images into fewer tokens. We design a caption-based evaluation system that leverages large language models (LLMs) to predict content complexity and determine the optimal compression ratio for a given image, taking into account factors critical to human perception. Trained on images with diverse compression ratios, CAT demonstrates robust performance in image reconstruction. We also utilize its variable-length latent representations to train Diffusion Transformers (DiTs) for ImageNet generation. By optimizing token allocation, CAT improves the FID score over fixed-ratio baselines trained with the same flops and boosts the inference throughput by 18.5%.