The field of diffusion models and multimodal generation is rapidly advancing, with a focus on improving efficiency, flexibility, and accuracy. Recent developments have led to the creation of more efficient serving systems, such as dynamic stage-level serving paradigms, which can significantly reduce latency and improve resource allocation. Additionally, decentralized training methods have been proposed, enabling the training of high-quality diffusion models without the need for centrally coordinated infrastructure. Multimodal generation has also seen significant advancements, with the development of models that can handle variable-length and concurrent mixed-modal generation, as well as models that can interpret human sketches and generate 3D flight paths for drone navigation. Asynchronous denoising diffusion models have also been proposed, which can improve text-to-image alignment by dynamically modulating the timestep schedules of individual pixels. Noteworthy papers include: TridentServe, which proposes a dynamic stage-level serving paradigm to improve the efficiency of diffusion pipelines. Paris, which presents a decentralized trained open-weight diffusion model that achieves high-quality text-to-image generation without centrally coordinated infrastructure. OneFlow, which enables concurrent mixed-modal generation and outperforms autoregressive baselines on both generation and understanding tasks. SketchPlan, which generates 3D flight paths for drone navigation from human sketches. Lumina-DiMOO, which introduces a fully discrete diffusion modeling approach for seamless multi-modal generation and understanding. DreamOmni2, which proposes multimodal instruction-based editing and generation tasks and achieves impressive results.