Capability-Driven Data for Generative AI

The advancement of large-scale image generation has long been constrained by the siloed optimization of task-specific datasets. A fundamental challenge lies not just in curating individual corpora, but in orchestrating diverse supervision signals according to the dependencies between generative capabilities. This is the problem addressed by a new capability-driven data infrastructure.

Orchestrating Relational Supervision Across Generative Tasks

This framework introduces a novel approach by coupling capability-specific supervision construction with capability-aligned curriculum scheduling. It employs three specialized, interoperable data engines. These engines are designed to build complementary relational supervision for crucial generative tasks: text-image grounding, inter-image transformation, and image-knowledge association. Complementing this, dedicated caption experts ensure alignment of text-to-image (T2I) and editing supervision across different tasks and granularities. This structured approach to multimodal diffusion models training data is key to overcoming previous limitations.

Curriculum Learning for Evolving Generative Competencies

A multi-stage curriculum is central to this infrastructure. It jointly evolves task composition, visual-concept distribution, data quality, and image resolution. This evolution follows the dependency order of capability acquisition, ensuring a logical progression of learning. The loop is closed through capability-aware evaluation, which uses targeted retrieval, expert construction, and gap-aware resampling to continuously refine the process. This methodology underpins the efficient creation of extensive multimodal diffusion models training data, including a 440M-image T2I corpus, 120M editing pairs, and over 27M image-entity pairs.

Scaling Generative Models with Structured Data

The practical impact of this infrastructure is demonstrated through the training of multimodal diffusion models at unprecedented scales. The researchers trained models of 3B and 6B sizes from scratch. Quantitative evaluations on CPI-Bench, alongside qualitative assessments across diverse text-to-image and editing scenarios, reveal broad visual coverage, versatile rendering capabilities, and effective transfer across generative abilities. This work provides a blueprint for building the next generation of generative AI models.