Research map / Image generation / Editing and style transfer
StableI2I: Spotting Unintended Changes in Image-to-Image Transition
Jiayang Li, Shuo Cao, Xiaohui Li, Zhizhen Zhang, Kaiwen Zhu, Yule Duan, Yu Qiao, Jian Zhang, Yihao Liu
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 |
|---|---|---|---|
| apply RL to the base model (ablation) | Qwen3-VL-8B-Instruct · Binary & Type QA average | accuracy: 33.17 -> 64.95 | Improved |
| add SFT before RL (ablation) | StableI2I RL · Binary & Type QA average | accuracy: 64.95 -> 80.47 | Improved |
| add multiple-choice QA in the RL stage only (ablation) | StableI2I SFT+RL · Binary & Type QA average | accuracy: 80.47 -> 80.73 | No clear effect |
| add multiple-choice QA in the SFT stage as well as RL (ablation) | StableI2I SFT+RL · Binary & Type QA average | accuracy: 80.73 -> 81.90 | Improved |
| add multiple-choice QA in both stages (ablation) | StableI2I SFT+RL · Binary & Type QA average | accuracy: 80.47 -> 81.90 | Improved |
| add multiple-choice QA in both stages (ablation) | StableI2I SFT+RL · Binary & Type QA low-level | accuracy: 94.10 -> 95.00 | Improved |
| add RL on augmented data (Dev.2 vs Dev.1) (ablation) | StableI2I-Dev.1 · Binary & Type QA semantic | accuracy: 59.00 -> 67.70 | Improved |
| add RL on augmented data (Dev.2 vs Dev.1) (ablation) | StableI2I-Dev.1 · Binary & Type QA structure | accuracy: 83.70 -> 80.30 | Hurt |
| re-train with augmented plus original data (final vs Dev.1) (ablation) | StableI2I-Dev.1 · StableI2I-Bench total | accuracy: 85.58 -> 87.35 | Improved |
| re-train with augmented plus original data (final vs Dev.2) (ablation) | StableI2I-Dev.2 · StableI2I-Bench total | accuracy: 86.35 -> 87.35 | Improved |
| add multi-stage training and data augmentation (ablation) | Qwen3-VL-8B-Instruct · StableI2I-Bench total | accuracy: 37.53 -> 87.35 | Improved |
| add multi-stage training on multiple-choice QA (ablation) | Qwen3-VL-8B-Instruct · Multiple-choice QA average | accuracy: 41.90 -> 91.70 | Improved |
Closest papers by meaning
- I2EBench: A Comprehensive Benchmark for Instruction-based Image Editing (NeurIPS 2024 · similarity 0.88)
- Who Evaluates the Evaluations? Objectively Scoring Text-to-Image Prompt Coherence Metrics with T2IScoreScore (TS2) (NeurIPS 2024 · similarity 0.88)
- Balancing Preservation and Modification: A Region and Semantic Aware Metric for Instruction-Based Image Editing (ICML 2025 · similarity 0.87)
- Arena-T2I Hard: Benchmarking and Improving Faithfulness with Dependency-Aware Checklist Rewards (NeurIPS 2026 · similarity 0.86)
- DynT2I-Eval: A Dynamic Evaluation Framework for Text-to-Image Models (NeurIPS 2026 · similarity 0.86)
- EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing (ICLR 2026 · 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.
