DIM: A Unified Multimodal Image Editing Model That Rebalances the Roles of Designer and Painter

DIM (Draw-In-Mind) is a study accepted by ICLR 2026, proposed by the ShowLab team at the National University of Singapore. By clearly distinguishing between the roles of "designer" (understanding design intent) and "painter" (executing painting operations), this model resolves the core contradiction of role confusion in existing unified multimodal models and significantly improves image editing capabilities.

Unified multimodal models are powerful in image tasks, but existing architectures often conflate "understanding design intent" and "executing painting operations", leading to either neglect of detailed execution or lack of overall control. This role confusion is the core contradiction in image editing, and the DIM framework brings a new breakthrough to address this.

Core Ideas

DIM draws on the division of labor in human creativity, separating the roles of designer (conceiving style and composition) and painter (visual presentation) to balance semantic understanding and pixel manipulation.

Technical Innovations

1. Dual-Path Representation Learning: Processes semantic-level design concepts and pixel-level visual details in parallel to avoid information compression loss;
2. Dynamic Role Switching: Adjusts role weights according to task requirements (designer leads for major changes, painter leads for fine adjustments);
3. Hierarchical Instruction Parsing: Identifies intent and details in editing instructions and routes them to the corresponding path.

DIM was accepted by ICLR 2026 and leads in multiple benchmark tests:

• Object replacement and insertion: Seamlessly integrates new objects with consistent backgrounds;
• Style transfer: Preserves content structure and accurately applies target styles;
• Attribute editing: Precisely controls visual attributes like color and texture;
• Composite editing: More stable handling of complex instructions. In addition, it is highly robust to ambiguous instructions and can proactively clarify and complete incomplete instructions.

• Creative Design Assistance: Quickly explore visual solutions via natural language descriptions;
• Content Creation Tools: Lower the threshold for professional image processing;
• Intelligent Image Restoration: Balance the coordination between restored areas and their surroundings;
• Multimodal Dialogue Systems: Provide a foundation for deep image editing dialogue assistants.

The ShowLab team open-sourced the DIM code and pre-trained models:

1. Provides a reference paradigm for role separation, which can be extended to other multimodal tasks;
2. Facilitates researchers to reproduce and improve, accelerating progress in the field;
3. Can be directly integrated into creative tool platforms by the industry.

• Introduce more professional roles (e.g., "critic", "curator");
• Design more efficient information exchange mechanisms between roles;
• Extend to multimodal content editing such as video and 3D;
• Learn user/domain role preferences to provide personalized experiences.