DataEvolver: Self-Improving Visual Data Engine via Goal-Driven Loop Agents
DataEvolver, an advanced visual data engine, has been developed to enhance the generation and management of visual data. It accommodates various data types, such as RGB images, depth maps, and meshes. The system integrates two primary operational loops: one for refining sample generation and another for expanding validation across datasets. Featuring a fixed Qwen-Edit LoRA probe, the updated Ours+DualGate model surpasses its predecessor in managing object rotation within images. This state-of-the-art framework is poised to improve controllability in visual data processing, aiding both in image refinement and multimodal comprehension applications.
Key facts
- DataEvolver is a closed-loop visual data engine.
- It automates iterative generation, inspection, correction, filtering, and export.
- Supports RGB images, masks, depth maps, normal maps, meshes, poses, trajectories, and review traces.
- Two coupled loops: generation-time self-correction and validation-time self-expansion.
- Validated on image-level object-rotation setting.
- Uses fixed Qwen-Edit LoRA probe.
- Ours+DualGate model outperforms unadapted base model.
- Addresses bottleneck in controllable visual data for image editing and multimodal understanding.
Entities
Institutions
- arXiv