Learning Generation Order Boosts Multimodal Diffusion Models

Recent advancements in Diffusion Language Models (DLMs) have shown promise in natural language generation, with adaptive token generation order improving performance in specific applications. This new work extends this concept to multimodal masked diffusion models, investigating how generation order impacts text-to-image synthesis and multimodal understanding. The researchers found that simple model logits are insufficient for determining optimal generation sequences in these complex multimodal tasks. To overcome this, they developed a learnable control module, trained using Group Relative Policy Optimization (GRPO), which dynamically determines the generation order. This novel approach demonstrably enhances both the alignment between text prompts and generated images, particularly in capturing fine-grained spatial details, and significantly improves the model's capabilities in multimodal reasoning and comprehension tasks.