TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation

TL;DR

TokenFlow introduces a dual-codebook image tokenizer that effectively separates semantic understanding from pixel-level generation, leading to significant improvements in multimodal understanding and high-quality image synthesis.

Contribution

It proposes a novel dual-codebook architecture that decouples semantic and pixel features, enhancing both understanding and generation capabilities in a unified framework.

Findings

Outperforms LLaVA-1.5 13B in understanding tasks with 7.2% improvement

Achieves FID score of 0.63 at 384x384 resolution for image reconstruction

Sets new state-of-the-art in autoregressive image generation with GenEval score of 0.55 at 256x256

Abstract

We present TokenFlow, a novel unified image tokenizer that bridges the long-standing gap between multimodal understanding and generation. Prior research attempt to employ a single reconstruction-targeted Vector Quantization (VQ) encoder for unifying these two tasks. We observe that understanding and generation require fundamentally different granularities of visual information. This leads to a critical trade-off, particularly compromising performance in multimodal understanding tasks. TokenFlow addresses this challenge through an innovative dual-codebook architecture that decouples semantic and pixel-level feature learning while maintaining their alignment via a shared mapping mechanism. This design enables direct access to both high-level semantic representations crucial for understanding tasks and fine-grained visual features essential for generation through shared indices. Our extensive experiments demonstrate TokenFlow's superiority across multiple dimensions. Leveraging TokenFlow, we demonstrate for the first time that discrete visual input can surpass LLaVA-1.5 13B in understanding performance, achieving a 7.2\% average improvement. For image reconstruction, we achieve a strong FID score of 0.63 at 384*384 resolution. Moreover, TokenFlow establishes state-of-the-art performance in autoregressive image generation with a GenEval score of 0.55 at 256*256 resolution, achieving comparable results to SDXL.