DreamOmni3: Scribble-based Editing and Generation
Published:Dec 27, 2025 09:07
•1 min read
•ArXiv
Analysis
This paper introduces DreamOmni3, a model for image editing and generation that leverages scribbles, text prompts, and images. It addresses the limitations of text-only prompts by incorporating user-drawn sketches for more precise control over edits. The paper's significance lies in its novel approach to data creation and framework design, particularly the joint input scheme that handles complex edits involving multiple inputs. The proposed benchmarks and public release of models and code are also important for advancing research in this area.
Key Takeaways
- •DreamOmni3 enables flexible image editing and generation using scribbles, text, and images.
- •It introduces a novel joint input scheme to handle complex edits.
- •The paper defines several scribble-based editing and generation tasks.
- •Comprehensive benchmarks are established to promote further research.
- •Models and code will be publicly released.
Reference
“DreamOmni3 proposes a joint input scheme that feeds both the original and scribbled source images into the model, using different colors to distinguish regions and simplify processing.”