FreeVideo

FlashML-orgApache-2.0活跃维护

在普通电脑上跑 MiniMax H3 视频生成,最低 8GB 显存加 16GB 内存就能用,还会根据硬件自动选择加速方案。

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.

v0.4.37发布于 2026-11-09 13:29 UTC2,915star +81 / 7 天362fork同步于 2026-11-10 05:41 UTC

怎么上手

由 README 提炼,执行前请确认
cd ComfyUI/custom_nodes
git clone https://github.com/FlashML-org/FreeVideo.git
硬件要求按 README 说明

最低 8 GB 显存 + 16 GB 内存;GeForce 与 RTX 30 系走 int8,工作站与数据中心卡走 FP8;Apple silicon Mac 为预览版,已在 24 GB 统一内存的 M5 上测试。

8G 显存可跑支持 Apple Silicon
下载:源码 zip·GitHub Releases(安装包)本站不镜像安装包

项目说明

由 README 自动翻译提炼

在自有电脑上本地跑 MiniMax H3 生成视频,最低 8 GB 显存即可

在普通电脑上跑 MiniMax H3 视频生成,最低 8GB 显存加 16GB 内存就能用,还会根据硬件自动选择加速方案。

核心能力

  • 按显卡自动选权重格式,GeForce 与 RTX 30 系用 int8、工作站卡用 FP8
  • 权重流式加载、异步预取与分块计算压低峰值显存
  • 支持文本、首尾帧以及图像、视频、音频参考输入
  • 作为 ComfyUI 插件提供创作工作区,支持两遍采样与批量生成
  • 启动器一键装好 ComfyUI、运行环境和模型,可复用已有模型

容易踩的坑

模型需单独下载,离线安装要四个模型包加环境包;Mac 预览版未公证,首次打开会被系统拦截需手动放行;作为 ComfyUI 自定义节点安装后需重启并在 Settings 里完成配置。

内容由模型根据 README 生成,可能不完整或不准确,以官方文档为准。看原文 README

最近发版

5 个版本v0.4.37
  1. v0.4.37FreeVideo v0.4.37最新正式版2026-11-09 · 2026-11-09 13:29 UTC

    Download / 下载

    Windows 10/11 · NVIDIA GPUFreeVideo.exe
    macOS 14+ · Apple siliconFreeVideo-Mac-arm64.dmg
    Linux x86_64 · NVIDIA GPUFreeVideo-Linux-x86_64.AppImage

    Windows: run FreeVideo.exe, choose an existing ComfyUI folder or install a new one, then click Install & launch. Offline packages are on Quark and ModelScope.

    macOS: open the DMG and drag FreeVideo.app into Applications. If macOS asks you to confirm the first launch, see the first-open steps.

    Linux: make the AppImage executable (chmod +x FreeVideo-Linux-x86_64.AppImage) and open it, as on Windows. On a server or from a terminal, see Linux.

    Windows: 运行 FreeVideo.exe,选择已有的 ComfyUI 或安装新的 ComfyUI,点击 安装并启动。离线包见夸克网盘或魔搭。

    macOS: 打开 DMG,将 FreeVideo.app 拖入「应用程序」。首次打开如需确认,请参考首次打开步骤。

    Linux: 给 AppImage 加上可执行权限(chmod +x FreeVideo-Linux-x86_64.AppImage)后打开,用法与 Windows 相同。在服务器或终端中使用请参考 Linux 说明。

    What's new

    Generation previews are a frame sheet again with an optional loop, the editor can narrate a line in a chosen voice, Replace character takes two people, and Windows installs download less.

    • The preview shown while a video is generated is a sheet of frames again. A new setting, Preview during generation, plays it as a loop of the video in progress instead. The preview is also shown when the window is short.
    • In the editor, right-click an empty spot on the dialogue or recording track and choose Narration: write a line, pick a voice, and it is generated to the length of the selection. New tutorial chapter and tour card.
    • Replace character takes a second person: the first picture replaces the person on the left, the second the person on the right (up to two people).
    • Generated music follows the written description and avoids an instrument it rules out, such as "no piano". The first-use download of the music files shows its percent on the generation bar.
    • Windows: when the graphics driver reports an error in the second pass, only the second pass is redone from the finished first pass instead of ending the request.
    • Windows installs download about 469 MB less: the ComfyUI workflow-template packages are no longer installed.
    • For multi-segment generation, segment progress now covers decoding and reaches 100% when finished.
    • Export and subtitle jobs show their own names and progress in the Agent activity list; the creation panel shows planning and adding counts.
    • Editor: the lines panel takes at most 40% of the editing area and scrolls inside, the five tracks stay visible, the resample menu is hidden while narrating, the audio button is now "Separate original sound", and characters, scenes and props sit in one row of equal squares.
    • Tutorials add Text and subtitles, Replace two people and Your own LoRA. Tour: "Take me there" opens the right view first, the media tour starts with first and last frames, and the
  2. v0.4.36FreeVideo v0.4.362026-11-08 · 2026-11-08 13:17 UTC
    查看更新说明

    Download / 下载

    Windows 10/11 · NVIDIA GPUFreeVideo.exe
    macOS 14+ · Apple siliconFreeVideo-Mac-arm64.dmg
    Linux x86_64 · NVIDIA GPUFreeVideo-Linux-x86_64.AppImage

    Windows: run FreeVideo.exe, choose an existing ComfyUI folder or install a new one, then click Install & launch. Offline packages are on Quark and ModelScope.

    macOS: open the DMG and drag FreeVideo.app into Applications. If macOS asks you to confirm the first launch, see the first-open steps.

    Linux: make the AppImage executable (chmod +x FreeVideo-Linux-x86_64.AppImage) and open it, as on Windows. On a server or from a terminal, see Linux.

    Windows: 运行 FreeVideo.exe,选择已有的 ComfyUI 或安装新的 ComfyUI,点击 安装并启动。离线包见夸克网盘或魔搭。

    macOS: 打开 DMG,将 FreeVideo.app 拖入「应用程序」。首次打开如需确认,请参考首次打开步骤。

    Linux: 给 AppImage 加上可执行权限(chmod +x FreeVideo-Linux-x86_64.AppImage)后打开,用法与 Windows 相同。在服务器或终端中使用请参考 Linux 说明。

    What's new

    Props join the cast library beside characters and scenes, the editor stays usable in a short window, and a refresh no longer clears text selected in the chat.

    • Cast library: props are a third kind beside characters and scenes. A prop can be uploaded or generated, and mentioned in a prompt. When a continued shot has room for fewer references, the character face is kept first, then the scene, then the prop.
    • Editor: in a short window the audio tracks and the add-audio button below the timeline are no longer cut off. The preview gives up height first, and the editor scrolls when that is not enough.
    • Chat: text selected in messages and cards is kept when the page refreshes, so it can be copied, and the reason for a failed send keeps its selection.
    • Launcher: when the engine stops unexpectedly, the launcher window and the page say when it stopped and offer to reconnect, and feedback and diagnostic reports include the latest exit record.

    更新内容

    道具加入角色库,与角色和场景并列;窗口较矮时编辑器仍可完整使用;聊天刷新时不再清掉已选中的文字。

    • 角色库:道具成为与角色、场景并列的第三类,可上传或生成,并在提示词中引用;接续镜头可用的参考位置不足时,依次保留角色脸、场景、道具。
    • 剪辑:窗口较矮时,时间线下方的声音轨和“+ 声音轨”按钮不再被裁掉;预览区先让出高度,仍不够时编辑器出现滚动条。
    • 聊天:页面刷新时,消息和任务卡中已选中的文字不再被清掉,可正常复制;发送失败的原因也保留选区。
    • 启动器:引擎意外停止时,启动器窗口和页面提示停止的时间并提供重新连接入口;反馈和诊断报告带上最近一次退出记录。
    a2a490a3ea7f11c5f23b5e75ec3ec713ac7ef279cd2ea35b6bc3f931ea449bc6  FreeVideo.exe
    13a05925e2fc7d5018d1ac884597e4b0ee7a8a53fa9c4ea84da0eb46e410d2dc  FreeVideo-Mac-arm64.dmg
    ad0cc5c222587e58599324bbb2ca0b567b929a61d81c77ccde935d293777aba6  FreeVideo-Linux-x86_64.AppImage
    

    Discord · QQ Group / QQ 群 · WeChat Group / 微信群

  3. v0.4.35FreeVideo v0.4.352026-11-07 · 2026-11-07 05:11 UTC
    查看更新说明

    Download / 下载

    Windows 10/11 · NVIDIA GPUFreeVideo.exe
    macOS 14+ · Apple siliconFreeVideo-Mac-arm64.dmg
    Linux x86_64 · NVIDIA GPUFreeVideo-Linux-x86_64.AppImage

    Windows: run FreeVideo.exe, choose an existing ComfyUI folder or install a new one, then click Install & launch. Offline packages are on Quark and ModelScope.

    macOS: open the DMG and drag FreeVideo.app into Applications. If macOS asks you to confirm the first launch, see the first-open steps.

    Linux: make the AppImage executable (chmod +x FreeVideo-Linux-x86_64.AppImage) and open it, as on Windows. On a server or from a terminal, see Linux.

    Windows: 运行 FreeVideo.exe,选择已有的 ComfyUI 或安装新的 ComfyUI,点击 安装并启动。离线包见夸克网盘或魔搭。

    macOS: 打开 DMG,将 FreeVideo.app 拖入「应用程序」。首次打开如需确认,请参考首次打开步骤。

    Linux: 给 AppImage 加上可执行权限(chmod +x FreeVideo-Linux-x86_64.AppImage)后打开,用法与 Windows 相同。在服务器或终端中使用请参考 Linux 说明。

    What's new

    External Agents can join the best part of every cut from different takes, the record button stays in place, and text and subtitles can be added after the sound has been separated.

    • External Agents: when the takes of a shot have the same cuts, an Agent can read each take's cuts, per-part checks and how the takes join at every cut, then join the best part of every cut from different takes. The choice is kept with the film.
    • External Agents: when a voice is changed with a script, a sentence split by a pause is said once instead of twice, and numbers, hyphenated words and initials stay whole.
    • Editor: the record button stays in the toolbar and turns grey when the selection cannot take a recording, so the icons after it no longer move.
    • Editor: audio track names fit in the track head, and text and subtitles can be added after the sound has been separated.
    • Tutorial and feature tour: new cards and chapters for Describe each part and Reuse settings. The example button reads "Use example" in both.

    更新内容

    外部 Agent 可从不同版本中选取每个切镜最好的一段并拼接,录音按钮位置固定,声音分离后可以添加文字和字幕。

    • 外部 Agent:同一镜头的各版本切镜相同时,Agent 可读取每一版的切镜、各段检查结果和各切镜处的衔接情况,从不同版本中为每个切镜选取最好的一段并拼接;所选片段随影片保存。
    • 外部 Agent:传入台词换人声时,被停顿切开的一句话只说一次;数字、连字符词和缩写保持完整。
    • 剪辑:录音按钮固定在工具栏中,选区不能录音时变灰,其后的图标不再移动。
    • 剪辑:声音轨名称在轨头内完整显示,声音分离后可以添加文字和字幕。
    • 教程和功能导览:新增“逐段描述”和“复用设置”的卡片和章节;示例按钮统一为“填入示例”。
    d89f5ccc2e13eb5e358f3ecd4dc79487c1002ee75caf9537d7da8f1835d93e1e  FreeVideo.exe
    752c68ab0e24e9690fcc904dd2b68e922a5d0b1a1025384508286c11c647aea7  FreeVideo-Mac-arm64.dmg
    39f8e671d3804b80bede62691ec80212e645e82ba5df2fbcab1451ed8ee08bfa  FreeVideo-Linux-x86_64.AppImage
    

    Discord · QQ Group / QQ 群 · WeChat Group / 微信群

  4. v0.4.34FreeVideo v0.4.342026-11-06 · 2026-11-06 07:59 UTC
    查看更新说明

    Download / 下载

    Windows 10/11 · NVIDIA GPUFreeVideo.exe
    macOS 14+ · Apple siliconFreeVideo-Mac-arm64.dmg
    Linux x86_64 · NVIDIA GPUFreeVideo-Linux-x86_64.AppImage

    Windows: run FreeVideo.exe, choose an existing ComfyUI folder or install a new one, then click Install & launch. Offline packages are on Quark and ModelScope.

    macOS: open the DMG and drag FreeVideo.app into Applications. If macOS asks you to confirm the first launch, see the first-open steps.

    Linux: make the AppImage executable (chmod +x FreeVideo-Linux-x86_64.AppImage) and open it, as on Windows. On a server or from a terminal, see Linux.

    Windows: 运行 FreeVideo.exe,选择已有的 ComfyUI 或安装新的 ComfyUI,点击 安装并启动。离线包见夸克网盘或魔搭。

    macOS: 打开 DMG,将 FreeVideo.app 拖入「应用程序」。首次打开如需确认,请参考首次打开步骤。

    Linux: 给 AppImage 加上可执行权限(chmod +x FreeVideo-Linux-x86_64.AppImage)后打开,用法与 Windows 相同。在服务器或终端中使用请参考 Linux 说明。

    What's new

    Agent can finish the sound of a whole editor project in one job, a line of a clip can be respoken in the original voice, and the Edit tab can save the current frame as a picture.

    • Agent: a new job finishes the sound of an editor project in one step. It separates each clip's sound, removes sustained music that came with the clips while keeping effects such as gunshots and crowds, lays one continuous music track across the cut, evens out loudness and fills silent gaps with room tone.
    • Edit only some of the lines in a clip: those lines are respoken in the original speaker's voice at the original timing, and the rest stay as they were. Dubbed lines keep the original pace and loudness. The dubbing model (about 4.9 GB) is downloaded on first use.
    • A reference video marked as voice only is now read as an audio reference: only its sound track is used and no picture is encoded. A voice reference used across several shots is no longer encoded again for each one.
    • Agent: take checks listen for the scripted lines (Chinese and English), notice a single odd frame at the start or end and a picture that stops before the take ends, and do not pick such takes. The Agent skill text also submits storyboards as one batch and resumes a retryable failure once.
    • Short dramas: long shots submitted right after startup are no longer refused while this computer's generation limit is still being worked out.
    • The Edit tab can save the frame under the playhead as a picture. It is drawn as the export would draw it and kept under Edit image > Generate with AI > History.
    • Continued parts in the editor can now use Preview first, and Run second pass replaces the preview in place.
    • The preview shown while a video is generated plays as a loop of up to sixteen frames instead of a still sheet of four, and no longer flashes black when its stage changes size. Recent videos and Creations fold videos with the same prompt together only when the
  5. v0.4.33FreeVideo v0.4.332026-11-05 · 2026-11-05 06:07 UTC
    查看更新说明

    Download / 下载

    Windows 10/11 · NVIDIA GPUFreeVideo.exe
    macOS 14+ · Apple siliconFreeVideo-Mac-arm64.dmg
    Linux x86_64 · NVIDIA GPUFreeVideo-Linux-x86_64.AppImage

    Windows: run FreeVideo.exe, choose an existing ComfyUI folder or install a new one, then click Install & launch. Offline packages are on Quark and ModelScope.

    macOS: open the DMG and drag FreeVideo.app into Applications. If macOS asks you to confirm the first launch, see the first-open steps.

    Linux: make the AppImage executable (chmod +x FreeVideo-Linux-x86_64.AppImage) and open it, as on Windows. On a server or from a terminal, see Linux.

    Windows: 运行 FreeVideo.exe,选择已有的 ComfyUI 或安装新的 ComfyUI,点击 安装并启动。离线包见夸克网盘或魔搭。

    macOS: 打开 DMG,将 FreeVideo.app 拖入「应用程序」。首次打开如需确认,请参考首次打开步骤。

    Linux: 给 AppImage 加上可执行权限(chmod +x FreeVideo-Linux-x86_64.AppImage)后打开,用法与 Windows 相同。在服务器或终端中使用请参考 Linux 说明。

    What's new

    Prompt enhancement works with less free graphics memory, and Setup or Repair now waits for a video that is still generating.

    • Prompts no longer receive the added sentence about natural sounds that 0.4.32 introduced.
    • Audio references: an mp4 file that only contains sound is accepted. An Agent that needs only the voice of a video can send it at a small picture size, which makes generation much faster.
    • Agent: choosing a take cancels the takes of that shot that are still queued or running.
    • Agent: when a voice reference is uploaded, the first six seconds are checked for speech, and the Agent is told when there is too little, such as in a clip that opens with music, so another clip can be chosen.
    • Setup or Repair that is stopped by a video still generating now waits for it and continues by itself, instead of showing a general failure.
    • Prompt enhancement: rewriting now works with less free graphics memory. When other programs hold the memory it needs, it waits and starts by itself once there is room, and lists those programs when their names can be read.
    • Audio processing in videos no longer differs depending on when FreeVideo was installed.
    • With several graphics cards, a new generation waits when its card is still briefly held by the previous request, instead of failing within a few seconds.
    • The Autopilot run bar stays fully visible in the top bar on narrow windows.
    • Settings: on narrow screens, the three Getting started buttons no longer break their labels across lines.

    更新内容

    提示词增强在可用显存较少时也能改写;安装或修复遇到还在生成的视频时会等待并自动继续。

    • 提示词不再被补写 0.4.32 加入的自然声音描述。
    • 音频参考:只有声音的 mp4 文件也可以使用。Agent 只需要视频里的声音时,可用小画面传入,生成明显更快。
    • Agent:选定一版后,同一镜头还在排队或生成的其他版本自动撤下。
    • Agent:上传声音参考时检查前 6 秒的说话时长,说话太少(例如以音乐开头)时告知 Agent,便于换一段。
    • 安装或修复被还在生成的视频拦下时,启动器现在会等视频结束后自动继续,不再显示通用的失败提示。
    • 提示词增强:可用显存较少时也能改写;显存被其他程序占用时会等待,空出后自动改写,能读取到程序名时一并列出。
    • 视频声音的处理结果不再因安装时间不同而有差异。
    • 多张显卡时,新的生成遇到显卡还被上一个请求短暂占用会等待,不再几秒内失败。
    • AI 全自动运行时,窄窗口顶栏的运行条完整显示。
    • 设置:窄屏

用户点评

登录用户的真实使用记录,可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译

    Make videos on the computer you already own. Powered by Video DeltaNet (VDN), 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

    News

    • 2026-10-13 · v0.4.0: AI agents, minute-long videos and an editing view. Let an AI agent such as Claude Code or Codex make videos on its own, generate videos over a minute long, and edit them on a timeline.
    • 2026-10-07 · v0.3.2: preview first. See a half-resolution preview first and finish only the videos you like.
    • 2026-10-07 · v0.3.0: a faster int8 model. GeForce and RTX 30-series cards switch to int8; an RTX 3090 makes a 5-second video about 2.3x faster.
    • 2026-10-06 · Gallery is live. Watch 20-second clips made with FreeVideo, and the four quality levels side by side.
    • 2026-10-06 · v0.2.3: videos carry their workflow. Drop a FreeVideo video onto the ComfyUI canvas to restore its prompt, seed and settings. Earlier videos can get theirs too.
    • 2026-10-05 · v0.2.0: four quality levels. Choose Light, Medium, High or Max for each video; higher levels give higher quality but take longer. Results can be exported as sharing images or videos with the generation time and GPU.
    • 2026-10-05 · v0.1.2: FreeVideo on Mac. Apple silicon Macs generate MiniMax H3 videos locally (preview).
    • 2026-10-03 · Community LoRAs. Use MiniMax H3 LoRAs right in your workflow. See examples.
    • 2026-10-02 · FreeVideo is open source. MiniMax H3 on consumer GPUs with as little as 8 GB of VRAM.

    About

    FreeVideo is a local inference engine for MiniMax H3 on consumer GPUs, built on OpenVDN's 8-step VDN-H3 model with Video DeltaNet'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 launchers for Windows, macOS and Linux. Its core features include:

    • Hardware-adaptive execution: Chooses the weight format for each GPU, int8 on GeForce and RTX 30-series cards and FP8 on workstation and datacenter cards, 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.
    • Multimodal inputs: Text prompts, first and last frames, and image, video and audio references.
    • Community LoRAs: Use MiniMax H3 LoRAs in your workflow. See examples.
    • ComfyUI integration: A dedicated creative workspace inside ComfyUI that supports two-pass sampling and batch generation and keeps a history of past creations. For finer control, switch to the node view to add LoRAs or customize the workflow.
    • One-click deployment: The Windows, Mac and Linux launchers set up ComfyUI, the runtime environment and the models, reuse existing models, and support offline installation.

    Getting Started

    Windows

    1. Download FreeVideo.exe and run it.
    2. Select an existing ComfyUI folder or install a new one. Existing model folders can be added for reuse; missing models are downloaded automatically.
    3. Click Install & launch. ComfyUI opens in the browser with the FreeVideo workspace.

    Security software: FreeVideo.exe is code-signed. If security software blocks FreeVideo.exe or quarantines files in the FreeVideo folder, check that the SHA-256 of FreeVideo.exe matches the value on the release page, then restore the files and add the FreeVideo folder to the software's exclusion list.

    Offline installation: Download the packages from Quark and drag the ZIP files into the launcher without extracting them. A fully offline installation needs the four model packs (video model, text encoder, video & audio decoder, and sampling caches), plus the environment package for a new ComfyUI installation. After importing only the environment package, you can also choose Automatic download for the models (v0.3.12 or newer). The audio reference cache is optional and is used only when a reference includes audio.

    macOS (Apple silicon preview)

    1. Download FreeVideo-Mac-arm64.dmg, open it and drag FreeVideo.app into Applications.
    2. Open FreeVideo, then select an existing ComfyUI folder or install a new one. Existing model folders can be added for reuse; the runtime environment and missing models are downloaded automatically.
    3. Click Install & launch. ComfyUI opens in the browser with the FreeVideo workspace.

    This preview has been tested on an M5 Mac with 24 GB of unified memory. See the Mac guide for generation times and memory.

    The Mac preview isn't notarized by Apple yet, so macOS blocks it the first time you open it. Download it only from the Releases page, then check the file and approve it as described in the Mac guide. This approves FreeVideo only; your other security settings stay as they are.

    Existing ComfyUI

    Install FreeVideo as a custom node:

    cd ComfyUI/custom_nodes
    git clone https://github.com/FlashML-org/FreeVideo.git
    

    Restart ComfyUI, open Workflow → Browse Templates → FreeVideo → FreeVideo-All-in-One, and complete the setup in FreeVideo Settings.

    Linux

    Desktop: Download FreeVideo-Linux-x86_64.AppImage, make it executable and open it, then follow the same steps as on Windows.

    Terminal or server:

    git clone https://github.com/FlashML-org/FreeVideo.git && cd FreeVideo
    ./setup.sh
    

    Setup installs FreeVideo, ComfyUI and the models, then opens FreeVideo in the browser. Afterwards, type freevideo to open it again. On a server, connect from your own computer through SSH, or run freevideo server --listen 0.0.0.0 for an access link on your local network. See the Linux guide.

    More details

    Support

    Report bugs in GitHub Issues, or ask questions on Discord, QQ or WeChat.

    Citation

    FreeVideo is based on VDN-H3. If you use FreeVideo in your research, please cite the Video DeltaNet paper:

    @article{xi2026videodeltanet,
      title={Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation},
      author={Xi, Haocheng and Xie, Yiming and Zhao, Hexu and Zhang, Yiwen and Liu, Michael and Creavin, Thomas and Keutzer, Kurt and Li, Xiuyu and Lv, Zhaoyang and Xu, Chenfeng and Feng, Haiwen},
      journal={arXiv preprint arXiv:2609.20744},
      year={2026}
    }
    

    Team

    Project Team

    Bowen Xue · Shuo Yang · Haocheng Xi · Xiaoze Fan · Chenfeng Xu

    Special Thanks

    Special thanks to AIwood爱屋研究室 and T8star-Aix for testing the project and providing valuable feedback.

    Listed in chronological order of participation.

    Acknowledgment

    We thank OpenVDN for Video DeltaNet / VDN-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 for the base model and the following projects: ComfyUI, Diffusers, SageAttention, the MiniMax H3 latent upscaler, the H3 text encoder for ComfyUI and Qt for Python.

    License

    The code is released under the Apache License 2.0. The model weights are licensed under the MiniMax H3 Community License, which includes territorial and acceptable-use restrictions.