The LynnReal-AI team released beta 0.1 of the unified video model LynnReal-Omni: a single open checkpoint covers text- and image-based generation, pose control, editing, restoration of degraded video, and streaming of long clips — tasks that previously required a separate model for each. The flagship 32B version runs in 4 denoiser steps, while the lightweight 27B-parameter Flash version runs in 3 steps, bringing an open video model closer than ever to real-time speed on a single H100. Weights are on Hugging Face along with an official ComfyUI port and ready-made workflows, so you can try the model today.

image
image

What happened

The LynnReal-AI team released beta 0.1 of the LynnReal-Omni video model. The flagship 32B-parameter checkpoint is built as a multimodal diffusion transformer based on the MiniMax H3 architecture, with Qwen3-VL fragments serving as the conditioning encoder. A single model covers t2v, i2v, body and hand pose control, structural control from 3D renders and game captures, omni-reference generation, style transfer, video editing, frame-by-frame restoration of degraded video, and streaming of long videos in 4 denoiser steps. A separate 27B LynnReal-Omni-Flash version runs in 3 steps with W8A8 quantization and a lightweight VAE decoder: on a single H100, generating and decoding 22 frames at 540p takes 843 ms for Standard and 377 ms for Flash, while a 5-second 1344×768 clip with native stereo audio is assembled in about 8.4 seconds on Flash. Weights are on Hugging Face: Standard BF16 takes 61.7 GiB, Flash INT8 — 37.0 GiB. The release is accompanied by an official ComfyUI port with the ComfyUI-LynnReal node pack and ready-made workflows, a technical report on arXiv (number 2609.15863), and the MSAVP benchmark for multi-shot audio-visual generation.

Context

A classic video stack is assembled from a chain of narrow models: each task needs its own checkpoint, data formats must be reconciled between stages, and each link must be maintained separately. LynnReal-Omni offers a different design: a single model directly accepts heterogeneous conditions — appearance references, 3D renders, game captures — and works with them without intermediate pipelines. This kind of multi-conditionality is exactly what agentic visual pipelines need, where a soft agent drives a task from idea to finished clip without switching between services. Open weights are essential here: the team's key claims, from speed to task coverage, can be independently verified without relying on vendor assurances.

Why this matters for the industry

For companies building video pipelines, consolidating the stack into a single checkpoint means fewer services to maintain, fewer handoffs between stages, and simpler deployment. An agent can pass heterogeneous conditions to a single model instead of a chain of narrow service calls — a direct step toward agentic visual workflows, where the visual part of automation is assembled within a single checkpoint. The Flash version's speed brings an open video model closer to real-time on a single H100 card: if independent measurements confirm the team's numbers, real-time video streaming stops being the prerogative of closed systems and becomes an engineering task for the open stack. The official ComfyUI port moves the pipeline from research scripts to a mass-market tool around which custom workflows immediately form. If the beta matures, rapid integration into the ComfyUI ecosystem, quantized versions, and adaptations for smaller VRAM are likely.

Why this matters for users

You can try the model the same day: just copy the ComfyUI-LynnReal node pack, workflows, and models from the comfyui/models directory into your ComfyUI installation. Standard (4 denoiser steps) has workflows for t2v, i2v, r2v, pose2v, and v2v; Flash (3 steps) has workflows for t2v, ti2v, and ref2v. The hard barrier is hardware: the release was tested on H100 80GB and RTX 4090 48GB, and a standard 24 GB of VRAM will not be enough. Then come the beta rough edges: the Flash path in ComfyUI does not yet work with clips longer than 11 seconds, the port itself is experimental and slower than the original scripts, and the repository remains an early beta with known limitations.

What is still unknown / limitations

The stated timings are a point measurement by the team: open weights make the numbers reproducible, but extrapolation to streaming long videos requires evidence of consistency over long sequences. The MSAVP benchmark was presented by the same team that released the model, and the methodology, baselines, and comparison protocols are not disclosed, so qualitative comparisons with other models should be considered self-eval until external audit, and quality assessments should be built on your own tests. This is an early beta 0.1: at the current pace of field development, specific version numbers will quickly become outdated, and independent verifications and competitor reactions have not yet occurred.

Sources

Author

Look at AI, editorial team