Research map / Vision-language and multimodal / Multimodal LLMs
What Changed? Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models
Lorenzo Baraldi, Davide Bucciarelli, Federico Betti, Marcella Cornia, Lorenzo Baraldi, Nicu Sebe, Rita Cucchiara
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 stage 1 pretraining on similar real image pairs (ablation) | DICE difference detector (Idefics3-8B) · DICE-D difference detection (class-agnostic) | AP: 15.1 -> 16.7 | Improved |
| add stage 1 pretraining on similar real image pairs (ablation) | DICE difference detector (Idefics3-8B) · DICE-D difference detection (class-aware) | AP: 9.4 -> 10.6 | Improved |
| add stage 2 inpainting fine-tuning after stage 1 (ablation) | DICE difference detector (Idefics3-8B) · DICE-D difference detection (class-agnostic) | AP: 2.9 -> 22.3 | Improved |
| add extracted command-probability confidence instead of fixed value 1 (ablation) | DICE difference detector (Idefics3-8B, both stages) · DICE-D difference detection (class-agnostic) | AP: 16.7 -> 22.3 | Improved |
| add extracted confidence instead of fixed value 1 (stage 2 only) (ablation) | DICE difference detector (Idefics3-8B) · DICE-D difference detection (class-agnostic) | AP: 15.1 -> 18.2 | Improved |
| use extracted confidence instead of random confidence (ablation) | DICE difference detector (Idefics3-8B) · DICE-D difference detection (class-agnostic) | AP: 16.8 -> 22.3 | Improved |
| add extracted confidence in the coherence AP computation (ablation) | DICE (Idefics3-8B) · DICE-D coherence estimation (detected areas) | AP: 12.0 -> 15.5 | Improved |
| use Idefics3-8B instead of mPLUG-Owl3-7B as backbone | DICE difference detector · DICE-D difference detection (class-agnostic) | AP: 5.2 -> 22.3 | Improved |
| use Idefics3-8B instead of Qwen2-VL-7B as backbone | DICE difference detector · DICE-D difference detection (class-agnostic) | AP: 2.3 -> 22.3 | Improved |
| use Idefics3-8B instead of mPLUG-Owl3-7B as coherence estimator | DICE coherence estimation · DICE-D coherence estimation (detected areas) | accuracy: 85.9 -> 85.4 | No clear effect |
| patch coherent differences in CLIP-I using DICE | CLIP-I · Correlation with human judgment, background preservation | correlation rho: 51.1 -> 54.5 | Improved |
| patch all detected differences in CLIP-I | CLIP-I · Correlation with human judgment, background preservation | correlation rho: 51.1 -> 32.7 | Hurt |
Closest papers by meaning
- EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing (ICLR 2026 · similarity 0.89)
- InsightEdit: Towards Better Instruction Following for Image Editing (CVPR 2025 · similarity 0.88)
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.87)
- SmartEdit: Exploring Complex Instruction-based Image Editing with Multimodal Large Language Models (CVPR 2024 · similarity 0.87)
- ReasonEdit: Towards Reasoning-Enhanced Image Editing Models (CVPR 2026 · similarity 0.87)
- The Promise of RL for Autoregressive Image Editing (NeurIPS 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.
