arXiv new; image tokenizer/autoregressive generation · P0 · 2026-07-01

GEAR：它和通用视觉自监督的关系在于：把 VQ tokenizer 与 AR generator 端到端联合训练，并用 representation alignment 让生成器反向塑造视觉 token

高相关；详见方法、贡献和实验边界。

GEAR: Guided End-to-End AutoRegression for Image Synthesis　arXiv new; image tokenizer/autoregressive generation　原文链接

编号2606.32039 优先级P0 类别arXiv new; image tokenizer/autoregressive generation 会议arXiv new; image tokenizer/autoregressive generation 方法把 VQ tokenizer 与 AR generator 端到端联合训练，并用 representation alignment 让生成器反向塑造视觉 token 来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：把 VQ tokenizer 与 AR generator 端到端联合训练，并用 representation alignment 让生成器反向塑造视觉 token。高相关；详见方法、贡献和实验边界。

**(c) GEAR (Ours) Figure 2 Overview of GEAR**(c) GEAR (Ours) Figure 2 Overview of GEAR. (a) The conventional pipeline freezes a pretrained VQ-VAE and trains the AR model alone with the next-token-prediction (NTP) and REPA losses. (b) Naively making the pipeline end-to-end by passing AR gradients back into the tokenizer through the straight-through estimator (STE, drawn as the zigzag arrows <sup>⇝</sup>) is highly unstable and collapses (cf. table 7). (c) GEAR reads the per-position assignment both as a hard (one-hot) and a soft (temperature-scaled) matrix: the hard branch carries NTP and the hard REPA loss to update only the AR model, while the diferentiable soft branch carries a REPA loss that bypasses the upper AR blocks and flows back (the dashed arrow <sup>99K</sup>) to update only the tokenizer, giving a stable end-to-end guidance signal.这张图概括 GEAR 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。

**(b) gFID across model sizes (w/o vs**(b) gFID across model sizes (w/o vs. w/ CFG) Figure 1 GEAR accelerates and improves autoregressive image generation. (a) gFID versus training steps on ImageNet (without CFG): GEAR converges up to 10× faster than LlamaGen-REPA, whereas the naive end-to-end variant that back-propagates into the tokenizer through the straight-through estimator diverges (gFID≈105). (b) gFID across model scales at 1.5M steps: GEAR improves performance at every size (B/L/XL), both without and with CFG.这张图/表用于判断 GEAR 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。

核心问题

它和通用视觉自监督的关系在于：把 VQ tokenizer 与 AR generator 端到端联合训练，并用 representation alignment 让生成器反向塑造视觉 token。

方法拆解

把 VQ tokenizer 与 AR generator 端到端联合训练，并用 representation alignment 让生成器反向塑造视觉 token

主要贡献

高相关；详见方法、贡献和实验边界。

实验看点

实验部分建议重点看两类证据：一是作者是否把方法收益和更强数据、更长训练、更大模型区分开；二是跨模型、跨数据或跨任务迁移是否还能保留同样趋势。

局限与读法

这篇论文的结论需要结合任务设置、训练数据规模和消融实验一起看；不要只凭单个指标判断它对通用视觉表征的价值。