CAT: Content-Adaptive Image Tokenization

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Shen, Junhong, Tirumala, Kushal, Yasunaga, Michihiro, Misra, Ishan, Zettlemoyer, Luke, Yu, Lili, Zhou, Chunting
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866916552649998336
author Shen, Junhong
Tirumala, Kushal
Yasunaga, Michihiro
Misra, Ishan
Zettlemoyer, Luke
Yu, Lili
Zhou, Chunting
author_facet Shen, Junhong
Tirumala, Kushal
Yasunaga, Michihiro
Misra, Ishan
Zettlemoyer, Luke
Yu, Lili
Zhou, Chunting
contents 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%.
format Preprint
id arxiv_https___arxiv_org_abs_2501_03120
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CAT: Content-Adaptive Image Tokenization
Shen, Junhong
Tirumala, Kushal
Yasunaga, Michihiro
Misra, Ishan
Zettlemoyer, Luke
Yu, Lili
Zhou, Chunting
Computer Vision and Pattern Recognition
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%.
title CAT: Content-Adaptive Image Tokenization
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2501.03120