Research map / Image generation / Editing and style transfer
PhysEdit: Physically Consistent Image Editing via Causal Enforcement
Siqi Wan, Jingwen Chen, Yehao Li, Yingwei Pan, Ting Yao, Tao Mei
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 causal DiT (cDiT) to base editor (ablation) | OmniGen2 · PICABench-Superficial | overall accuracy (%): 45.36 -> 52.23 | Improved |
| add causal DiT (cDiT) to base editor (ablation) | OmniGen2 · PICABench-Superficial LSE | accuracy (%): 48.52 -> 63.96 | Improved |
| add consistency loss (CL) on top of cDiT (ablation) | PhysEdit · PICABench-Superficial | overall accuracy and consistency: +1.17% Acc, +0.31 dB Con (52.23 -> 53.40, 26.45 -> 26.76) | Improved |
| add physics-aware RL (PARL) on top of cDiT+CL (ablation) | PhysEdit · PICABench-Superficial | overall accuracy (%): 53.40 -> 56.72 | Improved |
| add PARL (ablation) | PhysEdit · PICABench-Superficial Causality | accuracy (%): 46.12 -> 49.54 | Improved |
| increase trajectory length N from 1 to 4 (ablation) | Base + cDiT · PICABench-Superficial | accuracy and consistency: +3.29% Acc, +0.78 dB Con (48.94 -> 52.23) | Improved |
| increase trajectory length N from 4 to 6 (ablation) | Base + cDiT · PICABench-Superficial | overall accuracy (%): 52.23 -> 52.41 | No clear effect |
| fine-tune single-step base on static pairs from PhysEdit-50K (ablation) | OmniGen2 · PICABench-Superficial | overall accuracy (%): 45.36 -> 48.94 | Improved |
| train on dense trajectories (full PhysEdit) instead of static SFT pairs (ablation) | OmniGen2 · PICABench-Superficial | overall accuracy (%): 48.94 -> 56.72 | Improved |
| use PhysEdit framework on Qwen-Image-Edit backbone | Qwen-Image-Edit · PICABench-Superficial | overall accuracy (%): 57.21 -> 60.70 | Improved |
| use PhysEdit framework on Qwen-Image-Edit backbone | Qwen-Image-Edit · ImgEdit-Bench | overall score: 4.16 -> 4.29 | Improved |
| use PhysEdit framework on Qwen-Image-Edit backbone | Qwen-Image-Edit · RISEBench | overall accuracy (%): 8.9 -> 9.1 | Improved |
Closest papers by meaning
- PhyEditBench: A Real-World Multi-Stage Benchmark for Physics-Aware Image Editing (ECCV 2026 · similarity 0.92)
- From Statics to Dynamics: Physics-Aware Image Editing with Latent Transition Priors (ICML 2026 · similarity 0.92)
- Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing (NeurIPS 2026 · similarity 0.87)
- Learning Action and Reasoning-Centric Image Editing from Videos and Simulation (NeurIPS 2024 · similarity 0.87)
- PhyS-EdiT: Physics-aware Semantic Image Editing with Text Description (CVPR 2025 · similarity 0.87)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.86)
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.
