LynnReal-Omni Unifies Video Tasks in 32B DiT, 27B Flash at 377ms

The Flash variant of LynnReal-Omni renders a 22-frame 540p video in 377 milliseconds on a single H100, according to a new arXiv preprint from Xiaofeng Mao, Peijia Lin, Shaohao Rui, Yibo Zhang, Haibin Wan and Weijie Ma. The full 32B model runs the same clip in 843 milliseconds; the Flash version trims to 27B parameters while keeping the shared multimodal diffusion transformer architecture.

The pitch is one model that swallows heterogeneous visual inputs and does most video work end to end. The abstract frames the design as combining stochastic generation with "explicit references, editable 3D scenes, or executable game states for stable control." That single weight set is claimed to cover text-to-video, image conditioning, reference-guided generation, structural control, editing, video restoration, and extended-length content generation.

Alongside the model, the authors ship MSAVP, a 100-prompt evaluation with 20 metrics grouped into "instruction adherence, plausibility, visual quality, temporal behavior, and audio synchronization." It's a self-published benchmark; the abstract doesn't break results out per capability or say how the 20 metrics track human preference.

This lands in a busy month for omni-capable video systems on our tracker, where separate labs are folding audio, references and structural signals into one diffusion stack rather than chaining specialists.