Construction of a Universal Foundation Model for Image Generation

Establish a concrete and validated construction for a universal foundation model for image generation that unifies diverse image-generation tasks within a single framework, resolving the current lack of clarity about how such a model should be designed and trained.

Background

The paper surveys prior diffusion and multimodal generation approaches, noting that most focus primarily on text-to-image generation and rely on task-specific extensions that do not generalize to more complex image-generation tasks. In this context, the authors argue that despite recent progress, a truly universal foundation model for image generation—capable of handling a broad spectrum of tasks within one architecture—has not yet been clearly constructed or thoroughly explored.

OmniGen is proposed as an initial attempt toward such unification, simplifying the pipeline by jointly modeling text and images without separate encoders and demonstrating end-to-end handling of multiple image-generation tasks. Nevertheless, the authors explicitly acknowledge that the general construction of a universal foundation model remains unclear, highlighting an open direction for future research beyond their current instantiation.