LynnReal-Omni is a video generation framework from Lynnreal Lab (GitHub org LynnReal-AI) that unifies text-to-video, image-to-video, human- and hand-pose-guided generation, structural control, omni-reference generation, style transfer, video editing, degraded-video restoration and streaming long-video generation inside one 32B-parameter shared multimodal diffusion transformer built on the MiniMax H3 architecture, running Standard generation in four denoiser steps; a companion 27B LynnReal-Omni-Flash variant drops to three steps with a lighter VAE decoder, and the README reports a single H100 producing a warm 22-frame 540p clip in 843 ms with Standard versus 377 ms with Flash. The repository, created September 13, 2026, ships inference code, sampling launchers and a ComfyUI port with workflows for t2v, i2v, r2v, pose2v and v2v on Standard four-step checkpoints and t2v, ti2v and ref2v on the Flash three-step checkpoint; its ComfyUI-LynnReal node pack auto-selects the fastest available attention backend and tunes INT8 GEMMs per GPU. On a single H100 80GB at 1344x768, the README measures a warm 5-second Flash text-to-video clip at 8.4s generate / 12.2s click-to-video, and flags clips longer than 11 seconds as not yet usable on the accelerated ComfyUI Flash path. Lite checkpoint variants (Standard BF16 37.6GiB, Standard INT8 20.4GiB, Flash INT8 16.7GiB) are reported bit-exact to their full-size counterparts at the shipped step counts. The weights -- Standard BF16 DiT, Flash W8A8 DiT and a lightweight VAE -- are hosted separately on Hugging Face under stdstu123 and are not gated; a Standard DiT INT8 checkpoint is listed as coming soon. The model and weights are distributed under MiniMax's H3 Community License Agreement, a non-standard, territorially restricted license: it is non-transferable, requires separate authorization for commercial use above $20 million in annual revenue, and by its own clause excludes the European Union, United Kingdom, Republic of Korea and United States. The README labels the project an early beta with expected bugs and inconsistent quality, and lists training code and additional datasets as roadmap items not yet released.