Researchers propose a new data infrastructure for generalist image generation that focuses on organizing heterogeneous supervision according to the dependencies among generative capabilities. The framework consists of three specialized data engines that build relational supervision for text-image grounding, inter-image transformation, and image-knowledge association. It also includes a multi-stage curriculum that evolves task composition, visual-concept distribution, and image resolution along the dependency order of capability acquisition. The authors demonstrate their approach by training multimodal diffusion models on large-scale datasets and conducting quantitative and qualitative evaluations. --- Why it matters: This work matters to researchers in AI because it addresses a central challenge in generalist image generation: how to organize heterogeneous supervision according to the dependencies among generative capabilities. By providing a capability-driven data infrastructure, the authors make it easier for others to build on their work and advance the field. Source: https://arxiv.org/abs/2608.18076