Research map / Image generation / Editing and style transfer
Shifting the Breaking Point of Flow Matching for Multi-Instance Editing
Carmine Zaccagnino, Fabio Quattrini, Enis Simsar, Marta Tintore Gazulla, Rita Cucchiara, Alessio Tonioni, Silvia Cascianelli
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 instance-disentangled attention (IDAttn) (ablation) | FLUX.1 Kontext · LoMOE-Bench | Background LPIPS: 0.126 -> 0.091 | Improved |
| add instance-disentangled attention (IDAttn) (ablation) | FLUX.1 Kontext · LoMOE-Bench | Target CLIP score: 25.22 -> 25.67 | Improved |
| add efficient prompt encoding to IDAttn (ablation) | Ours · LoMOE-Bench | Target CLIP score: 25.67 -> 25.60 | Hurt |
| add efficient prompt encoding to IDAttn (ablation) | Ours · LoMOE-Bench | Background LPIPS: 0.091 -> 0.099 | Hurt |
| use harmonization masks in early and late layers with disentanglement in middle layers vs disentanglement everywhere (ablation) | Ours · LoMOE-Bench | Background LPIPS: 0.108 -> 0.091 | Improved |
| use harmonization in early and late layers vs harmonization in all layers (ablation) | Ours · LoMOE-Bench | Target CLIP score: 25.05 -> 25.67 | Improved |
| use disentangled masks in the middle layers instead of harmonized masks (ablation) | Ours · LoMOE-Bench | Attempt Rate: 80.21 -> 92.19 | Improved |
| add visual cues (boxes) to FLUX prompts | FLUX · LoMOE-Bench | Target CLIP score: 24.71 -> 24.49 | No clear effect |
| use multi-turn editing instead of single pass | FLUX · LoMOE-Bench | Target CLIP score: 24.71 -> 25.71 | Improved |
| fine-tune the method on Crello Edit training data | Ours · Crello Edit | CER: 0.61 -> 0.52 | Improved |
| fine-tune the method on Crello Edit training data | Ours · InfoEdit | CER: 0.64 -> 0.56 | Improved |
| fine-tune the method | Ours · Crello Edit | FID: 9.45 -> 10.85 | Hurt |
Closest papers by meaning
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing (NeurIPS 2025 · similarity 0.89)
- Unpaired Visual Editing with Self-Consistent Flow Matching (ICML 2026 · similarity 0.89)
- Constrained Look-ahead Guidance for Interference-Aware Flow Editing (NeurIPS 2026 · similarity 0.89)
- FlowDC: Flow-Based Decoupling-Decay for Complex Image Editing (CVPR 2026 · similarity 0.89)
- Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing (CVPR 2024 · similarity 0.87)
- Stable Flow: Vital Layers for Training-Free Image Editing (CVPR 2025 · 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.
