Research map / Image generation / Text-to-image diffusion
Re-Align: Structured Reasoning-guided Alignment for In-Context Image Generation and Editing
Runze He, Yiji Cheng, Tiankai Hang, Zhimin Li, Yu Xu, Zijin Yin, Shiyi Zhang, Wenxun Dai, Penghui Du, Ao Ma, Chunyu Wang, Qinglin Lu and 2 more
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 SFT on Re-Align-410K with IC-CoT (ablation) | BAGEL · OmniContext subset | Overall: 5.80 -> 6.77 | Improved |
| add reasoning-generation alignment (GRPO with CLIP surrogate reward) without diversity (ablation) | Re-Align SFT · OmniContext subset | Prompt Following: 7.51 -> 7.46 | No clear effect |
| add reasoning-generation alignment (GRPO) (ablation) | Re-Align SFT · OmniContext subset | CLIPout: 33.32 -> 33.50 | Improved |
| add reasoning-induced diversity to RGA (ablation) | Re-Align SFT+RGA · OmniContext subset | Overall: 6.80 -> 6.89 | Improved |
| add reasoning-induced diversity to RGA (ablation) | Re-Align SFT+RGA · OmniContext subset | Prompt Following: 7.46 -> 7.61 | Improved |
| use structured IC-CoT instead of no reasoning (ablation) | Re-Align · GSB human evaluation | win rate: 20% higher win rate | Improved |
| use structured IC-CoT instead of unstructured BAGEL-style CoT (ablation) | Re-Align · GSB human evaluation | win rate: 16.25% higher win rate | Improved |
| apply Re-Align training to BAGEL | BAGEL · OmniContext (average) | Overall score: 5.73 -> 8.21 | Improved |
| apply Re-Align training to BAGEL | BAGEL · DreamOmni2Bench Replace | Overall score: 1.37 -> 8.61 | Improved |
| compare Re-Align against Qwen-Image-Edit on single-subject generation | Re-Align · OmniContext SINGLE (Character) | Overall score: 8.35 -> 8.25 | Hurt |
| compare Re-Align against Emu3.5 and GPT-4o overall | Re-Align · OmniContext (average) | Overall score: 8.82/8.80 -> 8.21 | Hurt |
| Re-Align versus DreamOmni2 on Global editing | DreamOmni2 · DreamOmni2Bench Global | Overall score: 7.76 -> 7.85 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DreamOmni2 — DreamOmni2: Multimodal Instruction-based Generation and Editing (CVPR 2026) · better on 7 of 7 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
- GIR-Bench: Versatile Benchmark for Generating Images with Reasoning (ICLR 2026 · similarity 0.89)
- Factuality Matters: When Image Generation and Editing Meet Structured Visuals (ICLR 2026 · similarity 0.89)
- WEAVE: Unleashing and Benchmarking the In-context Interleaved Comprehension and Generation (CVPR 2026 · similarity 0.88)
- ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.88)
- MICON-Bench: Benchmarking and Enhancing Multi-Image Context Image Generation in Unified Multimodal Models (CVPR 2026 · similarity 0.88)
- ReFocus: Visual Editing as a Chain of Thought for Structured Image Understanding (ICML 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.
