论文Publications

单图参考扩散模型很难同时整合多个主体与明确空间布局。Single-reference diffusion models struggle to compose multiple entities under explicit spatial layouts.

无需训练，利用 Group Isolation Attention 和 Region-Modulated Attention 扩展 MMDiT 的多参考组合能力。Training-free multi-reference composition with Group Isolation Attention and Region-Modulated Attention on MMDiT.

在论文所用的多图组合基准上改善身份保持、背景一致性与区域布局控制。Reported gains in identity preservation, background consistency, and layout control on the evaluated composition benchmarks.

这些结果属于所测试的生成模型与基准；不能推断任意架构或复杂场景都同样有效。Findings depend on the evaluated models and benchmarks; they do not establish universal robustness.

图生视频生成会让图像取证痕迹随时间漂移，单帧检测难以追踪变化。Image-to-video synthesis moves forensic evidence over time, complicating frame-only detection.

把生成过程表述为像素随时间的运动，构建可学习的跟踪模板与模板引导的运动模块。Treats video generation as pixel motion over time using a learnable forensic template and template-guided flow.

论文报告了跨开放及商用 I2V 模型的时序取证泛化实验。The paper reports temporal-forensics results on commercial and open I2V generators.

目前属于公开预印本；实验结论应按文中数据与生成设置理解。This is a public preprint; conclusions are bounded by the evaluated models and settings.

生成图片需要同时证明来源与定位后续恶意编辑，单一稳健水印难以兼顾两者。Generated images need both provenance verification and post-edit tamper localization.

联合嵌入版权与定位水印，再通过扩散反演敏感性识别被篡改区域。Combines copyright/localization watermarking with diffusion inversion sensitivity for tamper localization.

论文在图像质量、抗失真与定位能力方面报告了多项实验。The ICCV paper reports robustness, localization, and image-quality evaluations.

水印的适用性需结合具体生成与编辑模型、威胁模型评估。Performance must be interpreted under the tested generators, edits, and threat assumptions.

面对投稿规模变化，需要可核对的学术数据了解研究热点与审稿生态。Changes in submission scale motivate an auditable overview of research trends and peer-review workload.

以 2026-10-08 的公开投稿快照为基础，开展主题分布、重复论文筛查及同行评审相关描述性分析。Descriptive analysis of the public 2026-10-08 ICLR 2027 submission snapshot, including topics and overlap audits.

研究页面报告覆盖 43,013 篇投稿的描述性统计。The analysis reports descriptive statistics over 43,013 submissions.

这是投稿分布的描述性研究，不等于最终录用趋势或论文质量的因果证据。Submission statistics cannot establish final acceptance patterns or causal paper-quality claims.