Lumina-DiMOO: An open-source discrete multimodal diffusion model (synbol.github.io)

Lumina-DiMOO is a newly introduced open-source foundational model that employs a fully discrete diffusion framework for multimodal generation and understanding, addressing challenges in seamlessly processing text, images, and other modalities. Unlike traditional autoregressive or hybrid AR-diffusion models, Lumina-DiMOO’s discrete diffusion approach enables more efficient sampling and supports a wide range of tasks such as text-to-image and image-to-image generation—including editing, subject-driven generation, and inpainting—while also excelling in image understanding. The significance of Lumina-DiMOO lies in its state-of-the-art performance across multiple benchmarks, surpassing leading open-source multimodal models despite using a moderate 8-billion parameter scale. Its discrete diffusion modeling notably enhances capabilities in complex attribute recognition, object counting, spatial reasoning, and detailed image generation, achieving superior scores in both generation and understanding tasks compared to models like DALL-E 3 and GPT-4o. By releasing both code and checkpoints, Lumina-DiMOO paves the way for researchers to explore discrete diffusion as a promising direction for unified multimodal AI, advancing the efficiency and versatility of next-generation multimodal systems.