Scholarlink

<p>Physically-consistent image editing remains challenging due to heterogeneous edit requirements and fixed inference schedules in existing methods. This work addresses spatial and dynamical heterogeneity in diffusion-based editing. We present PhysEdit, a region-aware framework with adaptive spatio-temporal reasoning. It includes <i>Complexity-Adaptive Reasoning Depth</i> (CARD) and <i>Spatial Reasoning Mask</i> (SRM) to allocate computation dynamically. Experiments on the ImgEdit dataset show 1.18× speedup over baselines while improving instruction adherence (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(+0.7\%\)</EquationSource> </InlineEquation> CLIP-T), with appearance-level edits reaching 1.52×. The approach maintains identity preservation and achieves category-dependent efficiency gains. This work provides a new paradigm for adaptive conditional computation in visual content editing. Code and usage instructions are publicly available at <a href="https://github.com/leeguandong/PhysEdit">https://github.com/leeguandong/PhysEdit</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive spatio-temporal reasoning for physically-consistent region-aware image editing

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Guandong Li,
  • Mengxia Ye

摘要

Physically-consistent image editing remains challenging due to heterogeneous edit requirements and fixed inference schedules in existing methods. This work addresses spatial and dynamical heterogeneity in diffusion-based editing. We present PhysEdit, a region-aware framework with adaptive spatio-temporal reasoning. It includes Complexity-Adaptive Reasoning Depth (CARD) and Spatial Reasoning Mask (SRM) to allocate computation dynamically. Experiments on the ImgEdit dataset show 1.18× speedup over baselines while improving instruction adherence ( \(+0.7\%\) CLIP-T), with appearance-level edits reaching 1.52×. The approach maintains identity preservation and achieves category-dependent efficiency gains. This work provides a new paradigm for adaptive conditional computation in visual content editing. Code and usage instructions are publicly available at https://github.com/leeguandong/PhysEdit.