Technology ID
TT-0036
Description
DreamOmni2 is an open-source model that unifies instruction-based image editing and generation. It processes text + multiple reference images to handle concrete objects and abstract concepts (e.g., style, mood, material, lighting). Key innovations include a three-step bootstrap data synthesis pipeline with feature mixing, index encoding for multi-reference handling, position encoding shift, and VLM joint training. This enables superior performance in complex tasks like object replacement, pose imitation, style transfer, and abstract attribute manipulation, outperforming many open-source and rivaling closed-source models.
Content
Innovatively proposed the task of multimodal instruction-based editing, significantly improving user experience
Surpasses closed-source image generation and editing models such as Nano-Banana
Innovative and scalable data collection method
Inventors
Applicable Industries
Potential Application
Technology Focus
TRL
Images
