From Corpora to Co-Evolving Capabilities: Capability-Centric Data Design for Generalist Image Generation
The paper proposes organizing image-generation training data around capability dependencies, not isolated task corpora.
Its system builds separate but connected data engines for text-image grounding, image editing, and image-entity association. The authors say it curates 440M text-to-image samples, 120M editing pairs, and more than 27M image-entity pairs. They use that infrastructure to train 3B- and 6B-parameter multimodal diffusion models from scratch, then evaluate them on CPI-Bench and qualitative generation/editing cases. ArXiv · AI/CL/LG's note
Its system builds separate but connected data engines for text-image grounding, image editing, and image-entity association. The authors say it curates 440M text-to-image samples, 120M editing pairs, and more than 27M image-entity pairs. They use that infrastructure to train 3B- and 6B-parameter multimodal diffusion models from scratch, then evaluate them on CPI-Bench and qualitative generation/editing cases. ArXiv · AI/CL/LG's note
score 5