genmoai/mochi

The best OSS video generation models, created by Genmo

What it solves

Mochi 1 is an open-source video generation model designed to bridge the gap between closed-source and open-source high-fidelity video generation. It focuses on high-fidelity motion and strong adherence to user prompts, specifically optimized for photorealistic styles.

How it works

It uses a 10 billion parameter diffusion model based on a novel Asymmetric Diffusion Transformer (AsymmDiT) architecture. This architecture streamlines text processing to focus more neural capacity on visual reasoning by using separate MLP layers and different hidden dimensions for text and visual streams. For compression, it employs an AsymmVAE that causally compresses videos to a 128x smaller size.

Who it’s for

Researchers and developers who want a high-capacity, hackable open-source video generation model that can be run on high-end GPUs (recommended 1 H100) or integrated via a composable API.

Highlights

  • AsymmDiT Architecture: A 10B parameter model that optimizes memory and processing by treating text and visual tokens asymmetrically.
  • LoRA Fine-tuning: Includes an easy-to-use trainer to customize the model on personal video datasets.
  • Open Weights: Released under a permissive Apache 2.0 license.
  • Flexible Interface: Provides a Gradio UI, a CLI, and a programmatic API for video generation.

Related

  • Project
    abhiTronix/vidgear

    VidGear is a high-performance, multi-threaded Python framework that simplifies complex video processing, streaming, and hardware-accelerated media tasks by wrapping various state-of-the-art libraries into a unified API.

  • Project
    Jakeschincariol/youtube-agent-skill

    A set of eleven Claude‑compatible “skills” plus six Python utilities that help YouTubers script, edit, optimise and audit videos—all run locally on plain Python 3 and never publish automatically.

  • Project
    SandAI-org/MAGI-1

    MAGI-1 is an autoregressive video generation model that produces high-fidelity videos chunk-by-chunk to ensure temporal consistency and physical accuracy.

  • Project
    samuelgursky/davinci-resolve-mcp

    An MCP server that lets AI assistants control DaVinci Resolve video editing software through its official scripting API, with full editing, grading, and rendering capabilities, plus an offline file-level server for cloud workflows.