OpenVDN/vdn-minimax-h3

VideoDeltaNet-H3: Live T2VA / I2VA / FL2VA / Ref2VA-like generation based on Minimax H3.

What it solves

Video generation models often suffer from slow inference speeds. VDN-H3 addresses this by enabling faster video generation—potentially faster than the video plays—while maintaining near-lossless visual quality and consistency.

How it works

VDN-H3 uses a hybrid-attention architecture based on the MiniMax H3 backbone. It introduces a highly efficient frame-wise linear attention branch alongside a softmax branch that preserves visual quality. This is implemented as a plug-and-play system using a separate linear attention branch and two small LoRA adapters that can be merged into the backbone during inference without altering the original weights.

Who it’s for

Developers and researchers working with video generation models who need high-performance inference on both data-center GPUs (like B200/H200) and consumer GPUs (via the FreeVideo integration).

Highlights

  • High-Speed Inference: Capable of generating a 14.4-second clip in approximately 9 seconds end-to-end on 8x B200 GPUs.
  • Versatile Generation Modes: Supports text-to-video (T2VA), image-to-video (I2VA), first-and-last-frame-to-video (FL2VA), and reference-image-to-video (Ref2VA-like) tasks.
  • Plug-and-Play Architecture: Uses LoRA adapters and a linear attention branch that can be merged into the backbone.
  • Open-Source Stack: Provides full access to weights, optimized inference code, and training recipes.

Related

  • Project
    FlashML-org/FreeVideo

    FreeVideo is an open‑source desktop app that runs the MiniMax H3 text‑to‑video model locally via a ComfyUI plugin. It adapts precision and streams weights so it can generate videos on consumer GPUs with as little as 8 GB VRAM, offering four quality levels, multimodal inputs, LoRA support, and AI‑agent‑driven workflows.

  • Project
    hao-ai-lab/FastVideo

    FastVideo is an open‑source framework for training, distilling, and running video diffusion models at real‑time speeds. It offers sparse‑attention back‑ends, LoRA finetuning, and a ready‑to‑use web app (Dreamverse) for streaming video generation, and runs on NVIDIA GPUs, Apple Silicon, and DGX Spark.

  • Project
    antirez/h3.c

    A native Metal inference engine for MiniMax-H3 on Apple Silicon, enabling local text-to-video and text-to-audio generation with advanced memory and performance optimizations.

  • Project
    Tencent-Hunyuan/HunyuanVideo-1.5

    A lightweight 8.3B parameter video generation model that enables high-quality text-to-video and image-to-video synthesis on consumer-grade GPUs.

  • Project
    PKU-YuanGroup/Helios

    Helios is a 14B parameter real-time long video generation model capable of producing minute-scale, high-quality videos at up to 19.5 FPS on a single H100 GPU.