Research map / Image generation / Editing and style transfer
PICABench: How Far are We from Physical Realistic Image Editing?
Yuandong Pu, Le Zhuo, Songhao Han, Jinbo Xing, Kaiwen Zhu, Shuo Cao, Bin Fu, Si Liu, Hongsheng Li, Yu Qiao, Wenlong Zhang, Xi Chen and 1 more
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 |
|---|---|---|---|
| fine-tune on PICA-100K (video-derived synthetic edits) | FLUX.1 Kontext · PICABench-Superficial | overall accuracy: 48.93 -> 50.64 | Improved |
| fine-tune on PICA-100K | Qwen-Image-Edit · PICABench-Superficial | overall accuracy: 58.29 -> 62.13 | Improved |
| fine-tune on PICA-100K | Qwen-Image-Edit · PICABench-Superficial | consistency (dB): 19.43 -> 21.91 | Improved |
| fine-tune on PICA-100K | FLUX.1 Kontext · PICABench-Superficial | consistency (dB): 24.57 -> 25.23 | Improved |
| fine-tune on PICA-100K | FLUX.1 Kontext and Qwen-Image-Edit · PICABench global state transition | accuracy: 48.79 -> 47.02 (Flux); 67.33 -> 65.54 (Qwen) | Hurt |
| train on real-video MIRA400K instead of synthetic PICA-100K (ablation) | FLUX.1 Kontext · PICABench-Superficial | overall accuracy: 50.64 -> 46.96 | Hurt |
| use intermediate prompts (with physical rationale) instead of superficial (ablation) | Bagel, Flux.1 Kontext, Qwen-Image-Edit · PICABench | overall accuracy: Bagel 45.07 -> 54.06 | Improved |
| use explicit prompts instead of superficial (ablation) | Bagel, Flux.1 Kontext, Qwen-Image-Edit · PICABench | overall accuracy: Qwen 58.29 -> 68.02 | Improved |
| use more detailed prompts (ablation) | Bagel, Flux.1 Kontext, Qwen-Image-Edit · PICABench | consistency (dB): Bagel 28.42 -> 15.20 | Hurt |
| use unified understanding+generation architecture instead of dedicated editing model | open-source models · PICABench-Superficial | overall accuracy | Hurt |
| add region annotations and a stronger VLM judge to PICAEval (ablation) | PICAEval · human Elo correlation | Pearson correlation | Improved |
Closest papers by meaning
- UltraEdit: Instruction-based Fine-Grained Image Editing at Scale (NeurIPS 2024 · similarity 0.85)
- Learning Action and Reasoning-Centric Image Editing from Videos and Simulation (NeurIPS 2024 · similarity 0.85)
- Do Image Editing Models Understand Lighting? (NeurIPS 2026 · similarity 0.84)
- ProEdit: Simple Progression is All You Need for High-Quality 3D Scene Editing (NeurIPS 2024 · similarity 0.84)
- PS-Diffusion: Photorealistic Subject-Driven Image Editing with Disentangled Control and Attention (CVPR 2025 · similarity 0.84)
- Creating Your Editable 3D Photorealistic Avatar with Tetrahedron-constrained Gaussian Splatting (CVPR 2025 · similarity 0.84)
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.
