Search papers, labs, and topics across Lattice.
Channel-wise Vector Quantization (CVQ) is introduced, replacing patch-wise image tokens with channel-wise tokens, quantizing each feature map channel instead of each patch feature vector. This allows representing images as discrete levels of visual details, which is then used in a Channel-wise Autoregressive (CAR) model that predicts image channels sequentially for text-to-image generation. Experiments show CVQ achieves 100% codebook utilization with a 16K+ codebook and CAR attains a DPG score of 86.7 and a GenEval score of 0.79, demonstrating strong text-to-image generation effectiveness.
Forget patch-based image tokenization: channel-wise quantization unlocks better codebook utilization and text-to-image generation by representing images as discrete levels of visual detail.
We present Channel-wise Vector Quantization (CVQ), a novel image tokenization paradigm that replaces patch-wise tokens with channel-wise tokens. Unlike conventional vector quantization, which assigns a discrete token to each patch feature vector, CVQ quantizes each channel of the feature map. This formulation represents an image as discrete levels of visual details, rather than as a grid of spatial patches. Based on CVQ, we introduce a new visual autoregressive framework with "next-channel prediction". Instead of rendering images patch by patch in raster order, our Channel-wise Autoregressive (CAR) model predicts image channels sequentially, producing progressively enriched visual details. Specifically, it first sketches global structure and then refines fine-grained attributes, akin to a human artist's workflow. Empirically, we show that: (1) CVQ achieves 100% codebook utilization with a 16K+ codebook size without any bells and whistles, and substantially improves reconstruction quality over conventional VQ; and (2) CAR attains a DPG score of 86.7 and a GenEval score of 0.79, demonstrating strong effectiveness for text-to-image generation.
