LynnReal-Omni: Open Source 3-Step Video Editing & Control

November 5, 2026

LynnReal-AI, LynnReal-Omni

Share this post:

LynnReal-Omni: Open Source 3-Step Video Editing & Control

LynnReal-Omni (LynnReal-AI/LynnReal-Omni) is an open source multimodal video generation, control, and editing framework built on the MiniMax H3 Diffusion Transformer architecture (arXiv:2609.15863). It unifies text-to-video, first frame conditioning, multiple subject reference generation, human body and hand pose control, 3D mesh and game render stylization, video appearance editing, and degraded frame restoration into a single model that runs in 4 steps (Standard 32B) or 3 steps (Flash 27B W8A8).

Studio availability. LynnReal-Omni is not available in AI FILMS Studio. This guide covers the open source research checkpoints and local deployment. In the Studio, use Kling O3 Video Edit, Runway Aleph, Wan 2.2 VACE, or Seedance 2.0 on the Video to Video page or in Studio Nodes.

LynnReal-Omni 3 step and 4 step multimodal video generation, control, and editing showcase

One Multimodal Backbone Across Nine Filmmaking Tasks

Rather than maintaining separate adapters for pose conditioning, identity reference, and video inpainting, LynnReal-Omni trains a shared multimodal Diffusion Transformer across heterogeneous visual and audio tokens.

Generation. Supports text-to-video, first frame image-to-video, synchronized speech and video synthesis, and streaming video continuation (with 2 optional refinement steps per subsequent chunk).
Structural and Reference Control. Accepts single subject and multiple subject identity references, full body and articulated hand pose guidance, and direct rendering from untextured 3D mesh animations or game engine captures.
Editing and Error Repair. Performs instruction guided appearance editing and frame by frame video restoration that repairs compression artifacts and removes accumulated autoregressive drift in long clips.

LynnReal-Omni first frame and multi reference conditioning input

First frame and multiple subject reference inputs lock character wardrobe and facial structure across camera cuts. Credit: LynnReal-AI.

LynnReal-Omni stylized concept and structural rendering input

Heterogeneous conditioning lets directors combine concept art, 3D blockouts, and motion trajectories in one pass. Credit: LynnReal-AI.

377ms Rendering With LynnReal-Omni-Flash

To hit interactive rendering speeds, the authors distilled a 27B LynnReal-Omni-Flash variant (stdstu123/LynnReal-Onmi-flash-beta-0.1) that uses 42 transformer blocks, spatial token selection, INT8 projection weights, and INT8 activations (W8A8), paired with a lightweight VAE codec (stdstu123/LynnReal-Onmi-light-vae).

On a single NVIDIA H100 GPU, warm generation and decoding of a 22 frame 540p video clip completes in 843 ms with the 4 step Standard model and 377 ms with the 3 step Flash model. For workstations with tighter memory budgets, the Flash Lite (16.7 GiB INT8 DiT) and Standard Lite (20.4 GiB INT8 or 37.6 GiB BF16) checkpoints reduce disk and VRAM footprints while preserving the 3 step and 4 step sampling schedules.

Checkpoints, Footprint & Hardware Requirements

Model Variant	Hugging Face Repository	Parameters & Precision	Sampling Steps	DiT Footprint	Warm Latency (22f 540p, 1x H100)
LynnReal-Omni Standard	`stdstu123/LynnReal-Onmi-beta-0.1`	32B (`BF16` / `INT8` Lite)	4 steps	37.6 GiB (`BF16` Lite) / 20.4 GiB (`INT8` Lite)	843 ms
LynnReal-Omni Flash	`stdstu123/LynnReal-Onmi-flash-beta-0.1`	27B (`W8A8` 42 block)	3 steps	16.7 GiB (`INT8` Flash Lite)	377 ms
Lightweight VAE Codec	`stdstu123/LynnReal-Onmi-light-vae`	Fast latent decoder	1 pass	Companion codec	Included in 377 ms total

Downloading Checkpoints & Directory Layout

Clone the repository and place the downloaded Hugging Face bundles inside the weight/ directory so the launcher locates the DiT transformer, text encoder, and lightweight VAE automatically.

git clone https://github.com/LynnReal-AI/LynnReal-Omni.git
cd LynnReal-Omni

# Download the 3 step Flash W8A8 bundle and Lightweight VAE
huggingface-cli download stdstu123/LynnReal-Onmi-flash-beta-0.1 --local-dir ./weight/flash
huggingface-cli download stdstu123/LynnReal-Onmi-light-vae --local-dir ./weight/light-vae

# Or download the 4 step Standard bundle for maximum visual fidelity
huggingface-cli download stdstu123/LynnReal-Onmi-beta-0.1 --local-dir ./weight/standard

Pass --weights weight/flash or --weights weight/standard to the inference launcher depending on whether your priority is 3 step interactive speed (16.7 GiB DiT) or 4 step reference fidelity. To compare other fast video editors on the same hardware tier, see our guides to JoyAI-Video-Edit for 24 FPS streaming editing, SAMA for instruction guided video editing, and FreeVideo for running H3 models on 8 GB VRAM.

Multi Reference Video Editing in AI FILMS Studio

LynnReal-Omni runs on local GPU servers. To combine character references, structural video controls, and instruction guided editing in the cloud, open the video workspace to run Kling O3 Video Edit, Runway Aleph, Wan 2.2 VACE, and Seedance 2.0, or build custom pipelines in the Nodes Graph Editor.