Researchers have introduced LynnReal-Omni, a native multimodal video generation framework designed to bridge two complementary approaches: stochastic video diffusion models that are difficult to control, and agentic visual creation that offers stable control but does not by itself guarantee high object or character fidelity.
According to the paper, LynnReal-Omni is built on a 32B shared multimodal diffusion transformer that unifies text-to-video, image-conditioned generation, reference-guided generation, structural control, editing, degraded video restoration, and long-video generation. It accepts heterogeneous visual inputs — including appearance references, editable 3D renders, and game recordings — allowing agents to compose visual conditions within a single model.
The authors also trained a dedicated 27B Flash shared multimodal diffusion transformer for real-time rendering. To support training, they built a systematic data pipeline covering video cleaning, subject association, multimodal annotation, and aligned control construction, yielding a curated corpus of multi-shot audiovisual segments.
For evaluation, the team introduces MSAVP, a 100-prompt, 20-metric design that separates instruction following, generation plausibility, visual quality, temporal behavior, and audio coordination.
On performance, the paper reports that LynnReal-Omni-Flash reduces inference cost through model and decoding acceleration, including a lightweight VAE decoder. On a single H100, warm generation and decoding of a 22-frame 540p video takes 843 ms with LynnReal-Omni and 377 ms with Flash. The authors position these results as a foundation for real-time streaming video generation and as a unified, controllable, and efficient basis for agentic visual creation.