LynnReal-Omni
A unified video diffusion model with 32 billion parameters based on the MiniMax H3 architecture. It supports multiple tasks including text-to-video, image-to-video, pose-guided generation, style transfer, and video editing. The model employs a four-step fast generation process and has been released as open source.
Timeline (last 90 days)
LynnReal-Omni is a 32B-parameter unified video diffusion model built on the MiniMax H3 architecture, consolidating text-to-video, image-to-video, pose guidance, structural control, style transfer, video editing, and video restoration into a single framework.
LynnReal-Omni has released its code, model weights, and ComfyUI nodes on GitHub (LynnReal-AI/LynnReal-Omni) and Hugging Face.
LynnReal-Omni employs a 32B-parameter shared multimodal diffusion Transformer following the MiniMax H3 architecture.
LynnReal-Omni adopts four-step fast generation, significantly compressing inference costs compared to traditional diffusion models that require dozens of sampling steps.
LynnReal-Omni officially claims to be able to repair video quality degradation caused by accumulated error.
LynnReal-Omni can accept multiple heterogeneous inputs, including appearance references, editable 3D rendering results, and game screen recordings.
On a single H100 in warm state, the standard version (32B, 4-step) takes 843 milliseconds to generate and decode a 22-frame 540p video.
Whether the quality of each sub-task in a multi-task unified model can reach the standard of dedicated models still needs to be verified by the community on real-world material.
All Facts (8)
Whether the quality of each sub-task in a multi-task unified model can reach the standard of dedicated models still needs to be verified by the community on real-world material.
50%UnverifiedLynnReal-Omni can accept multiple heterogeneous inputs, including appearance references, editable 3D rendering results, and game screen recordings.
50%UnverifiedOn a single H100 in warm state, the standard version (32B, 4-step) takes 843 milliseconds to generate and decode a 22-frame 540p video.
50%UnverifiedLynnReal-Omni employs a 32B-parameter shared multimodal diffusion Transformer following the MiniMax H3 architecture.
50%UnverifiedLynnReal-Omni adopts four-step fast generation, significantly compressing inference costs compared to traditional diffusion models that require dozens of sampling steps.
50%UnverifiedLynnReal-Omni officially claims to be able to repair video quality degradation caused by accumulated error.
50%UnverifiedLynnReal-Omni has released its code, model weights, and ComfyUI nodes on GitHub (LynnReal-AI/LynnReal-Omni) and Hugging Face.
50%UnverifiedLynnReal-Omni is a 32B-parameter unified video diffusion model built on the MiniMax H3 architecture, consolidating text-to-video, image-to-video, pose guidance, structural control, style transfer, video editing, and video restoration into a single framework.
50%