通用视觉自监督研究报
A Daily Digest of Visual Self-Supervised Learning
2026-06-01 图像表征 · VFM · JEPA · 视频预训练
project · P0 · 2026-06-01

RayDer：它和通用视觉自监督的关系在于：把真实视频动态因素当成可吸收扰动，用单一 transformer 扩展自监督 NVS，并报告清晰 scaling 行为

高相关；详见方法、贡献和实验边界。

RayDer: Scalable Self-Supervised Novel View Synthesis from Real-World Video　arXiv + project　原文链接

编号2605.31535 优先级P0 类别project 会议arXiv + project 方法把真实视频动态因素当成可吸收扰动，用单一 transformer 扩展自监督 NVS，并报告清晰 scaling 行为 来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：把真实视频动态因素当成可吸收扰动，用单一 transformer 扩展自监督 NVS，并报告清晰 scaling 行为。 高相关；详见方法、贡献和实验边界。

(a) Camera Estimation same model (b) Novel View Synthesis Figure 8: Final Architecture Ove
(a) Camera Estimation same model (b) Novel View Synthesis Figure 8: Final Architecture Ove(a) Camera Estimation same model (b) Novel View Synthesis Figure 8: Final Architecture Overview. RayDer unifies camera estimation (a) and novel view synthesis (b) in a single transformer backbone. Lightweight local intra-frame encoder and decoder layers handle high-resolution processing.这张图概括 RayDer 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。
Figureure 13 · : Learned Camera Geometry Scales with Data, Model Size, and Compute
Figureure 13 · : Learned Camera Geometry Scales with Data, Model Size, and ComputeFigure 13: Learned Camera Geometry Scales with Data, Model Size, and Compute. We track the four continuous camera pose errors – rotation and translation, each read out both via a probe on the camera tokens (RayZer [28] protocol; top rows) and via cross-scene transfer (XFactor [46] protocol; bottom rows) – as a function of training compute, evaluated zero-shot on DL3DV-10K [40]. Left: all errors decrease consistently with training data scale. Right: all errors decrease with model scale, with insufficient data again imposing a strong ceiling. Notably, there is no significant saturation at scale, indicating that further scaling will likely be beneficial.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 RayDer 的方法或实验，请结合正文精读段落一起看。

核心问题

它和通用视觉自监督的关系在于：把真实视频动态因素当成可吸收扰动，用单一 transformer 扩展自监督 NVS，并报告清晰 scaling 行为。

方法拆解

把真实视频动态因素当成可吸收扰动，用单一 transformer 扩展自监督 NVS，并报告清晰 scaling 行为

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。