TokBench: Evaluating Your Visual Tokenizer before Visual Generation

TL;DR

This paper introduces TokBench, a lightweight benchmark for evaluating the reconstruction quality of visual tokenizers and VAEs on text and face images, revealing their limitations in preserving fine details.

Contribution

The paper presents a novel, efficient benchmark for assessing visual tokenizer performance on challenging content, highlighting their shortcomings in fine-grained feature preservation.

Findings

Modern visual tokenizers struggle with fine-grained features at small scales.

Traditional metrics do not accurately reflect reconstruction quality for faces and text.

The benchmark is lightweight, requiring only 2GB memory and 4 minutes to run.

Abstract

In this work, we reveal the limitations of visual tokenizers and VAEs in preserving fine-grained features, and propose a benchmark to evaluate reconstruction performance for two challenging visual contents: text and face. Visual tokenizers and VAEs have significantly advanced visual generation and multimodal modeling by providing more efficient compressed or quantized image representations. However, while helping production models reduce computational burdens, the information loss from image compression fundamentally limits the upper bound of visual generation quality. To evaluate this upper bound, we focus on assessing reconstructed text and facial features since they typically: 1) exist at smaller scales, 2) contain dense and rich textures, 3) are prone to collapse, and 4) are highly sensitive to human vision. We first collect and curate a diverse set of clear text and face images from existing datasets. Unlike approaches using VLM models, we employ established OCR and face recognition models for evaluation, ensuring accuracy while maintaining an exceptionally lightweight assessment process <span style="font-weight: bold; color: rgb(214, 21, 21);">requiring just 2GB memory and 4 minutes</span> to complete. Using our benchmark, we analyze text and face reconstruction quality across various scales for different image tokenizers and VAEs. Our results show modern visual tokenizers still struggle to preserve fine-grained features, especially at smaller scales. We further extend this evaluation framework to video, conducting comprehensive analysis of video tokenizers. Additionally, we demonstrate that traditional metrics fail to accurately reflect reconstruction performance for faces and text, while our proposed metrics serve as an effective complement.