Research map / Vision-language and multimodal / Multimodal LLMs
VLTR: Vision-Language Tool Reasoning for Instruction-Guided Image Editing
Yike Wang, Yitao Yu, Shaohua Sun, Ping Luo
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 instruction decomposition into primitives (ablation) | VLTR random baseline · PIE-Bench++ | Structure Distance: 142.3 -> 76.6 | Improved |
| add instruction decomposition into primitives (ablation) | VLTR random baseline · PIE-Bench++ | VLM Rank: 5.2 -> 3.8 | Improved |
| add inline verifier (ablation) | VLTR V2 · PIE-Bench++ | Structure Distance: 76.6 -> 65.2 | Improved |
| add inline verifier (ablation) | VLTR V2 · PIE-Bench++ | VLM Rank: 3.8 -> 3.2 | Improved |
| add inline verifier (ablation) | VLTR V2 · PIE-Bench++ | CLIP-T: 0.650 -> 0.651 | No clear effect |
| add Bayes-UCB router and uncertainty-aware scheduler (ablation) | VLTR V3 · PIE-Bench++ | Structure Distance: 65.2 -> 28.5 | Improved |
| add Bayes-UCB router and uncertainty-aware scheduler (ablation) | VLTR V3 · PIE-Bench++ | ImageReward: 0.521 -> 0.758 | Improved |
| use heteroscedastic verifier instead of quality-only verification (ablation) | VLTR · Verifier calibration (400 cases) | FAR: 17.3% -> 7.3% | Improved |
| use heteroscedastic verifier instead of quality-only verification (ablation) | VLTR · Verifier calibration (400 cases) | ECE: 0.288 -> 0.142 | Improved |
| use full scheduler instead of quality-only scheduler (ablation) | VLTR · 100 difficult PIE-Bench++ cases | success rate: 63% -> 82% | Improved |
| raise quality threshold tau_q from 0.6 to 0.7 (ablation) | VLTR scheduler · PIE-Bench++ | success rate: 82% -> 79% | Hurt |
| set inter-tier disagreement weight lambda to 0 instead of 0.5 (ablation) | VLTR verifier · Verifier calibration (400 cases) | ECE: 0.142 -> 0.183 | Hurt |
Closest papers by meaning
- ReasonEdit: Towards Reasoning-Enhanced Image Editing Models (CVPR 2026 · similarity 0.90)
- Reasoning to Edit: Hypothetical Instruction-Based Image Editing with Visual Reasoning (ICML 2026 · similarity 0.90)
- 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 0.89)
- SmartEdit: Exploring Complex Instruction-based Image Editing with Multimodal Large Language Models (CVPR 2024 · similarity 0.88)
- Guiding Instruction-based Image Editing via Multimodal Large Language Models (ICLR 2024 · similarity 0.88)
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.
