Make videos on the computer you already own. FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, and adapts its acceleration path to your hardware.
docs: credit Video DeltaNet and research support (#8)
Haocheng Xieb4c69ddbf95
diff --git a/README.md b/README.mdindex 487f320..7cefca4 100644--- a/README.md+++ b/README.md@@ -11,13 +11,15 @@<p align="center">English · <a href="README.zh-CN.md">中文</a></p>-Make videos on the computer you already own. FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, and adapts its acceleration path to your hardware.+Make videos on the computer you already own. Powered by [Video DeltaNet (VDN)](https://openvdn.github.io/), FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, with acceleration adapted to your hardware.https://github.com/user-attachments/assets/ecda7d0d-7fbe-4e0c-8c29-8f3315bafc15## About-FreeVideo is a local inference engine for running MiniMax H3 on consumer GPUs. Built on OpenVDN's 8-step [VDN-H3](https://huggingface.co/OpenVDN/vdn-minimax-h3) model, it schedules VRAM, system memory and disk as a single memory hierarchy, and plans weight placement, compute precision and attention kernels according to the GPU architecture and available resources. FreeVideo is delivered as a ComfyUI plugin: the Windows launcher installs and starts ComfyUI, and on Linux FreeVideo can also be used from the command line. Its core features include:+FreeVideo is a local inference engine for MiniMax H3 on consumer GPUs, built on [OpenVDN](https://github.com/OpenVDN)'s 8-step [VDN-H3](https://huggingface.co/OpenVDN/vdn-minimax-h3) model with [Video DeltaNet](https://openvdn.github.io/)'s hybrid attention.++It coordinates VRAM, system memory and disk, adapting weight placement, compute precision and attention kernels to the available hardware. FreeVideo runs as a ComfyUI plugin, with a Windows launcher for setup and command-line support on Linux. Its core features include:- **Hardware-adaptive execution**: Chooses the FP8 compute path for each GPU architecture, either native FP8 or FP8 storage with BF16 compute, and automatically probes the available attention kernels.- **Low-memory inference**: Weight streaming, asynchronous prefetching and chunked computation keep peak memory low, enabling inference with as little as 8 GB of VRAM and 16 GB of RAM.@@ -101,7 +103,11 @@ Special thanks to [**AIwood爱屋研究室**](https://space.bilibili.com/5039340## Acknowledgment-FreeVideo is built on [MiniMax H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) and [VDN-H3](https://github.com/OpenVDN/vdn-minimax-h3), and uses the following projects:+We thank [OpenVDN](https://github.com/OpenVDN) for [Video DeltaNet / VDN-H3](https://github.com/OpenVDN/vdn-minimax-h3) and its open-source model weights, training code and inference implementation.++We thank Impossible Research for providing computation resources.++We also thank [MiniMax H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) for the base model and the following projects:[ComfyUI](https://github.com/Comfy-Org/ComfyUI),[Diffusers](https://github.com/huggingface/diffusers),[SageAttention](https://github.com/thu-ml/SageAttention),diff --git a/README.zh-CN.md b/README.zh-CN.mdindex c5ac2da..83e74da 100644--- a/README.zh-CN.md+++ b/README.zh-CN.md@@ -11,13 +11,15 @@<p align="center"><a href="README.md">English</a> · 中文</p>-FreeVideo 让 MiniMax H3 能够在消费级显卡上本地运行,最低只需 8GB 显存和 16GB 内存,并会根据硬件配置自动选择合适的加速路径。+FreeVideo 由 [Video DeltaNet(VDN)](https://openvdn.github.io/) 驱动,让 MiniMax H3 能够在消费级显卡上本地运行,最低只需 8GB 显存和 16GB 内存,并会根据硬件配置自动选择合适的加速路径。https://github.com/user-attachments/assets/ecda7d0d-7fbe-4e0c-8c29-8f3315bafc15## 简介-FreeVideo 是在消费级显卡上运行 MiniMax H3 的本地推理引擎。它基于 OpenVDN 的 8 步模型 [VDN-H3](https://huggingface.co/OpenVDN/vdn-minimax-h3),将显存、内存与磁盘作为统一的分级存储进行调度,并根据显卡架构与可用资源规划权重放置、计算精度和注意力内核。FreeVideo 以 ComfyUI 插件的形式提供:Windows 启动器用于安装和启动 ComfyUI,Linux 上也可以通过命令行使用。主要特性包括:+FreeVideo 是面向消费级显卡的 MiniMax H3 本地推理引擎,基于 [OpenVDN](https://github.com/OpenVDN) 的 8 步模型 [VDN-H3](https://huggingface.co/OpenVDN/vdn-minimax-h3) 和 [Video DeltaNet](https://openvdn.github.io/) 的混合注意力。++它统一调度显存、内存与磁盘,并根据硬件条件调整权重放置、计算精度和注意力内核。FreeVideo 以 ComfyUI 插件形式提供,配备 Windows 启动器,也支持 Linux 命令行。主要特性包括:- **硬件自适应**:针对不同显卡架构选择 FP8 计算路径(原生 FP8,或 FP8 存储配合 BF16 计算),并自动探测可用的注意力内核,无需手动配置。- **低显存推理**:通过权重流式加载、异步预取与分块计算降低峰值显存,最低只需 8GB 显存和 16GB 内存。@@ -101,7 +103,11 @@ FreeVideo 基于 VDN-H3。如在研究中使用 FreeVideo,请引用 [Video Del## 致谢-FreeVideo 基于 [MiniMax H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) 和 [VDN-H3](https://github.com/OpenVDN/vdn-minimax-h3) 构建,并使用了以下项目:+感谢 [OpenVDN](https://github.com/OpenVDN) 开源 [Video DeltaNet / VDN-H3](https://github.com/OpenVDN/vdn-minimax-h3) 的模型权重、训练代码与推理实现。++感谢 Impossible Research 提供计算资源。++同时感谢 [MiniMax H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) 提供基础模型,以及以下项目:[ComfyUI](https://github.com/Comfy-Org/ComfyUI)、[Diffusers](https://github.com/huggingface/diffusers)、[SageAttention](https://github.com/thu-ml/SageAttention)、