Research map / Image generation / Text-to-image diffusion
Draw-In-Mind: Rebalancing Designer-Painter Roles in Unified Multimodal Models Benefits Image Editing
Ziyun Zeng, David Junhao Zhang, Wei Li, Mike Zheng Shou
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 |
|---|---|---|---|
| give the editor an external CoT designer (GPT-4o) instead of the raw prompt (ablation) | DIM-4.6B-Edit · ImgEdit | Overall: 3.10 -> 3.67 | Improved |
| use other open MLLMs as external designer instead of GPT-4o (ablation) | DIM-4.6B-Edit · ImgEdit | Overall: 3.67 (GPT-4o) vs 3.43 to 3.59 | Mixed |
| add stage-2 fine-tuning on ShareGPT-4o-Image-CoT vs non-CoT ShareGPT-4o-Image (ablation) | DIM-4.6B-Edit · ImgEdit | Overall: 3.22 -> 3.34 | Improved |
| stage-2 tuning on UltraEdit-160K-CoT alone (ablation) | DIM-4.6B-Edit · ImgEdit | Overall: 2.76 -> 2.79 | No clear effect |
| combine all three CoT sources (full DIM-Edit) in stage 2 (ablation) | DIM-4.6B-Edit · ImgEdit | Overall: 3.34 -> 3.67 | Improved |
| add Global Layout Perception step to CoT (ablation) | DIM-Edit CoT · ImgEdit | Overall: 3.49 -> 3.67 | Improved |
| add Local Object Perception step to CoT (ablation) | DIM-Edit CoT · ImgEdit | Overall: 3.39 -> 3.67 | Improved |
| add Edit Area Localization step to CoT (ablation) | DIM-Edit CoT · ImgEdit | Overall: 3.42 -> 3.67 | Improved |
| add Edited Image Imagination step to CoT (ablation) | DIM-Edit CoT · ImgEdit | Overall: 3.35 -> 3.67 | Improved |
| initialize editing model from DIM-4.6B-T2I (trained on DIM-T2I) instead of from scratch (ablation) | DIM-4.6B-Edit · ImgEdit | Overall: 2.66 -> 3.67 | Improved |
| add ShareGPT-4o-Image to UltraEdit-4M in stage 1 (ablation) | DIM-4.6B-Edit · ImgEdit | Overall: 2.76 -> 3.22 | Improved |
| feed DIM-style CoT blueprints as instructions to a different model | Step1X-Edit · ImgEdit | Overall: 3.06 -> 2.70 | Hurt |
Closest papers by meaning
- Factuality Matters: When Image Generation and Editing Meet Structured Visuals (ICLR 2026 · similarity 0.90)
- WEAVE: Unleashing and Benchmarking the In-context Interleaved Comprehension and Generation (CVPR 2026 · similarity 0.89)
- Breaking Dual Bottlenecks: Evolving Unified Multimodal Models into Self-Adaptive Interleaved Visual Reasoners (ICML 2026 · similarity 0.89)
- ReasonEdit: Towards Reasoning-Enhanced Image Editing Models (CVPR 2026 · similarity 0.88)
- EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing (ICLR 2026 · similarity 0.88)
- Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (CVPR 2026 · 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.
