Research map / Image generation / Text-to-image diffusion
The Devil is in Attention Sharing: Improving Complex Non-rigid Image Editing Faithfulness via Attention Synergy
Zhuo Chen, Fanyue Wei, Runze Xu, Jingjing Li, Lixin Duan, Angela Yao, Wen Li
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 attention sharing with RoPE (w=1.0) in position-insensitive blocks (ablation) | FLUX default · Non-Rigid Editing Benchmark (200 prompt pairs) | Gemini-2.5-Pro score: 2.3467 -> 4.2317 | Improved |
| remove RoPE from attention sharing (w=0.0) (ablation) | attention sharing w/ RoPE · Non-Rigid Editing Benchmark (200 prompt pairs) | Gemini-2.5-Pro score: 4.2317 -> 4.7850 | Improved |
| remove RoPE from attention sharing (w=0.0) (ablation) | attention sharing w/ RoPE · Non-Rigid Editing Benchmark (200 prompt pairs) | CLIPimg: 0.9180 -> 0.8963 | Hurt |
| use adaptive position weight w from editing measurement (Mmin 0.9, Mmax 1.0) (ablation) | attention sharing w/o RoPE · Non-Rigid Editing Benchmark (200 prompt pairs) | Gemini-2.5-Pro score: 4.7850 -> 5.4250 | Improved |
| use adaptive position weight w from editing measurement (Mmin 0.9, Mmax 1.0) (ablation) | attention sharing w/ RoPE (FreeFlux) · Non-Rigid Editing Benchmark (200 prompt pairs) | GPT-5 score: 5.9650 -> 6.6567 | Improved |
| tighten thresholds from Mmin=0.8 to Mmin=0.9 (ablation) | SynPS · Non-Rigid Editing Benchmark (200 prompt pairs) | Gemini-2.5-Pro score: 5.2133 -> 5.4250 | Improved |
| use SynPS instead of FreeFlux | FreeFlux · PIE-Bench ChangePose (40 prompt pairs, 5 seeds) | Gemini-2.5-Pro score: 3.2417 -> 4.1700 | Improved |
| use SynPS instead of CharaConsist | CharaConsist · PIE-Bench ChangePose (40 prompt pairs, 5 seeds) | Gemini-2.5-Pro score: 3.4933 -> 4.1700 | Improved |
| use SynPS instead of FreeFlux | FreeFlux · Non-Rigid Editing Benchmark (200 prompt pairs) | CLIPimg: 0.9180 -> 0.9051 | Hurt |
| use SynPS instead of FreeFlux | FreeFlux · Non-Rigid Editing Benchmark (200 prompt pairs) | CLIPtxt: 0.2291 -> 0.2344 | Improved |
| use SynPS instead of StableFlow | StableFlow · PIE-Bench ChangePose (40 prompt pairs, 5 seeds) | CLIPimg: 0.9712 -> 0.9415 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- StableFlow — Stable Flow: Vital Layers for Training-Free Image Editing (CVPR 2025) · better on 6 of 6 shared results
- CharaConsist — CharaConsist: Fine-Grained Consistent Character Generation (ICCV 2025) · better on 6 of 6 shared results
- FlowEdit — FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models (ICCV 2025) · better on 6 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Improving Editability in Image Generation with Layer-wise Memory (CVPR 2025 · similarity 0.87)
- Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing (CVPR 2024 · similarity 0.87)
- PairEdit: Learning Semantic Variations for Exemplar-based Image Editing (NeurIPS 2025 · similarity 0.87)
- RelationAdapter: Learning and Transferring Visual Relation with Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
- EditBridge: Towards Faithful and Efficient Ultra-High-Resolution Image Editing (NeurIPS 2026 · similarity 0.86)
- Concept Lancet: Image Editing with Compositional Representation Transplant (CVPR 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.
