Research map / Image generation / Editing and style transfer
From Statics to Dynamics: Physics-Aware Image Editing with Latent Transition Priors
Liangbing Zhao, Le Zhuo, Sayak Paul, Hongsheng Li, Mohamed Elhoseiny
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add physically-grounded reasoning (frozen Qwen2.5-VL traces) (ablation) | Qwen-Image-Edit + SFT · PICABench | Mechanics: 53.43 -> 57.43 | Improved |
| add implicit visual thinking (transition queries) only (ablation) | Qwen-Image-Edit · PICABench | Optics: 66.33 -> 68.05 | Improved |
| add implicit visual thinking without textual reasoning (ablation) | Qwen-Image-Edit + SFT · PICABench | Mechanics: 53.43 -> 52.38 | Hurt |
| plain SFT on PhysicTran38K without reasoning or queries (ablation) | Qwen-Image-Edit · PICABench | Overall: 61.26 -> 61.79 | No clear effect |
| combine reasoning and visual thinking (full PhysicEdit) (ablation) | Qwen-Image-Edit · PICABench | Overall: 61.26 -> 64.86 | Improved |
| use DINOv2 and VAE targets together instead of DINO only (ablation) | PhysicEdit · PICABench | Overall: 63.00 -> 64.86 | Improved |
| use DINOv2 and VAE targets together instead of VAE only (ablation) | PhysicEdit · PICABench | Overall: 63.05 -> 64.86 | Improved |
| smooth timestep-aware modulation instead of hard switching between DINO and VAE (ablation) | PhysicEdit · PICABench | Overall: 63.52 -> 64.86 | Improved |
| dual DINO+VAE guidance instead of DINO only (ablation) | PhysicEdit · PICABench | Global State Transition: 70.16 -> 67.67 | Hurt |
| apply PhysicEdit on top of Qwen-Image-Edit | Qwen-Image-Edit · KRISBench | Overall Score: 65.56 -> 72.16 | Improved |
| apply PhysicEdit on top of Qwen-Image-Edit | Qwen-Image-Edit · PICABench-Superficial | Overall: 61.26 -> 64.86 | Improved |
Closest papers by meaning
- PhyEditBench: A Real-World Multi-Stage Benchmark for Physics-Aware Image Editing (ECCV 2026 · similarity 0.93)
- PhysEdit: Physically Consistent Image Editing via Causal Enforcement (ECCV 2026 · similarity 0.92)
- PhyS-EdiT: Physics-aware Semantic Image Editing with Text Description (CVPR 2025 · similarity 0.91)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.88)
- Beyond Static Vision: Scene Dynamic Field Unlocks Intuitive Physics Understanding in Multi-modal Large Language Models (ICLR 2026 · similarity 0.88)
- Learning Action and Reasoning-Centric Image Editing from Videos and Simulation (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
