Duix-Avatar(前 HeyGem):本地离线数字人视频生成深度拆解

Tech-Experiment #数字人#视频生成#本地推理#Docker#开源拆解
🇨🇳 中文

开源仅供学习:本文所涉项目均来自公开仓库,分析仅供技术研究。如需商业使用,请仔细阅读对应许可证条款。


项目背景

duixcom/Duix-Avatar(GitHub:github.com/duixcom/Duix-Avatar)——之前叫做 HeyGem,在 v1.0.5(2025 年 8 月)正式改名。由新加坡公司 Duix 出品,定位是「本地全离线数字人视频生成工作台」:上传一段真人视频,克隆外貌和声音,再输入文字脚本,离线生成口播视频。

截至拆解时:15,585 stars,2,652 forks,开放问题 422 个。仓库最后一次 push:2026-04-21,近 5 个月无代码更新。


核心功能三件套

Duix-Avatar 的架构由三个 Docker 服务组成:

服务镜像来源功能
guiji2025/fun-asr阿里 FunASR语音识别(ASR)
guiji2025/fish-speech-zimingFish-Speech语音合成(TTS)
guiji2025/duix.avatar自研(C 语言)视频合成引擎

三步流程:外貌克隆 → 声音克隆 → 唇形同步视频生成,全部通过本地 REST API 串联(训练接口 /v1/train,视频合成 127.0.0.1:8383/easy/submit)。


硬件门槛

这是第一个拦路虎。官方明确要求:

  • GPU:NVIDIA RTX 4070(必须是 NVIDIA,无 AMD/Apple Silicon 支持)
  • 内存:32GB RAM
  • 存储:130GB+(C 盘 100GB+,数据盘 30GB+)
  • Docker 镜像下载量:约 70GB,安装耗时约 30 分钟

更棘手的是,RTX 5070 / 5090(Blackwell 架构)不受支持。Issue #624 显示新卡 CUDA kernel 直接报错,Issue #601/#617 也有类似反映,目前官方无回应。如果你刚换了最新一代 NVIDIA 卡,大概率跑不起来。


「10 秒克隆」说法核查

用户流传最广的描述是「上传 10 秒视频即可克隆」。我在官方 README、文档站(docs.duix.com)和所有 Release Notes 里找不到任何对应的具体数字。

README 对训练时长没有量化描述。这个说法可能来自第三方评测视频或早期营销内容,在官方一手资料中暂无实证,引用时请注明存疑。


「本地离线」的真实含义

关于「全离线」——视频生成本身确实全离线,所有推理在本地 Docker 容器内跑,官方承诺无需联网。

但 README 里有一句非常关键的自白:

本地版:「Usable effect」(可用效果) 云 API 版:「Stunning and higher definition effect」(惊艳高清效果)

这是官方原文的刻意措辞,相当于亲口承认本地效果比云端差一档。Duix 的商业模式是:开源版作引流,核心变现靠 duix.com 云 API 付费套餐。

另外,实时对话功能完全需要访问 duix.com 云服务,不能本地运行。


许可证陷阱:自定义「DUIX.COM Community License」

这是最需要警惕的部分。许可证名称叫「DUIX.COM Community License」,不是任何 SPDX 标准开源许可证(SPDX 字段标注的是 NOASSERTION)。核心条款:

  1. 1,000 MAU 付费门槛:产品月活达到 1,000 用户就必须申请商业授权,授权与否由 DUIX.COM 单方面决定
  2. 强制品牌展示:网站、界面、文档均须显示「Built with DUIX.COM」
  3. IP 诉讼终止条款:对 DUIX.COM 提起任何知识产权诉讼,许可即刻终止
  4. 永久营销权:DUIX.COM 获得使用你的实现案例做推广宣传的永久免费权利

还有一个未解决的合规问题:仓库打包分发了 GPL-3.0 的 FFmpeg(含 x264/x265),却套自定义许可证,存在 GPL 合规冲突(Issue #596,目前无官方回应)。


与 Duix-Mobile 的区别

同一组织下还有 duixcom/Duix-Mobile(8,255 stars,C++ 语言),但两者定位完全不同:

Duix-AvatarDuix-Mobile
定位视频生成工作台实时交互 Avatar SDK
延迟离线批量生成<120ms 实时
平台Windows / UbuntuiOS/Android/车机/VR/IoT
用途输出视频文件嵌入 App 运行时

如果目标是在手机 App 里跑实时数字人,Duix-Mobile 才是对的仓库。


支持语言

英语、日语、韩语、普通话、法语、德语、阿拉伯语、西班牙语(8 种)。


关键数字汇总

指标数值
Stars15,585
Docker 镜像大小~70GB
最低 GPUNVIDIA RTX 4070
最低内存32GB RAM
最低存储130GB+
支持语言数8 种
最后更新2026-04-21
MAU 商业授权门槛1,000 用户

综合判断

Duix-Avatar 打包了成熟的开源组件(FunASR + Fish-Speech)加自研 C 视频引擎,工程落地度够高,本地离线也是真的。但四点硬伤明显:

  1. 硬件门槛极高:RTX 4070 + 32GB RAM,70GB 镜像,不是随手能跑的工具
  2. 新显卡不支持:Blackwell(RTX 50 系)目前全线坑,5 个月没有维护迹象
  3. 许可证不是开源:1,000 MAU 就触发商业授权条款,GPL 冲突未解决
  4. 本地效果差一档:官方自承,如果追求质量还是要付费用云端

适合人群:有闲置 NVIDIA GPU 工作站、想离线预研数字人技术的开发者。不适合:生产部署、移动端、或者对质量有严格要求的场景。


开源仅供学习,商业使用请仔细核查许可证条款。


🇬🇧 English

Duix-Avatar (formerly HeyGem): Local Offline Digital Human Video Cloning — Deep Teardown

Open source for learning only: All projects discussed are from public repositories, analysis is for technical research purposes only.


Project Background

duixcom/Duix-Avatar (GitHub: github.com/duixcom/Duix-Avatar) — formerly called HeyGem, officially renamed at v1.0.5 (August 2025). Built by Singapore-based company Duix, positioned as a “fully local offline digital human video generation workbench”: upload a real video, clone appearance and voice, input a text script, and generate talking-head video offline.

At teardown time: 15,585 stars, 2,652 forks, 422 open issues. Last push: 2026-04-21 — nearly 5 months without a code update.


Core Three-Component Architecture

Three Docker services form the backbone:

Service ImageSourceFunction
guiji2025/fun-asrAlibaba FunASRSpeech recognition (ASR)
guiji2025/fish-speech-zimingFish-SpeechText-to-speech (TTS)
guiji2025/duix.avatarProprietary (C language)Video synthesis engine

Three-step pipeline: appearance cloning → voice cloning → lip-sync video generation, chained via local REST APIs (training: /v1/train, video synthesis: 127.0.0.1:8383/easy/submit).


Hardware Requirements

The first barrier. Official requirements:

  • GPU: NVIDIA RTX 4070 (NVIDIA only — no AMD or Apple Silicon)
  • RAM: 32GB
  • Storage: 130GB+ (100GB+ on C: drive, 30GB+ data drive)
  • Docker image download: ~70GB, ~30 minutes to install

Worse: RTX 5070 / 5090 (Blackwell architecture) are not supported. Issue #624 shows new cards throw CUDA kernel errors immediately. Issues #601/#617 report the same. No official response as of this writing.


”10-Second Clone” Claim Check

The widely circulated claim is “upload a 10-second video to clone.” I found no corresponding number anywhere in official README, docs.duix.com, or Release Notes.

The README has no quantified training time. This figure likely originated from third-party review videos or early marketing content — unverified in official first-party sources.


What “Local Offline” Actually Means

Video generation itself is genuinely fully offline — all inference runs in local Docker containers, no network required.

But the README contains a telling self-assessment:

Local version: “Usable effect” Cloud API version: “Stunning and higher definition effect”

That’s official copy explicitly acknowledging local quality is one tier below cloud. Duix’s business model: open-source as lead generation, monetize via duix.com cloud API.

Real-time conversation requires duix.com cloud services — not runnable locally.


License Trap: Custom “DUIX.COM Community License”

The license is named “DUIX.COM Community License” — not any SPDX-standard open-source license (SPDX field shows NOASSERTION). Key clauses:

  1. 1,000 MAU commercial trigger: Once your product reaches 1,000 monthly active users, you must apply for commercial licensing — DUIX.COM decides whether to grant it
  2. Mandatory branding: Website, UI, and documentation must display “Built with DUIX.COM”
  3. IP litigation termination: Any IP lawsuit against DUIX.COM immediately terminates your license
  4. Perpetual marketing rights: DUIX.COM gets perpetual, royalty-free rights to use your implementation as a promotional case study

Unresolved compliance issue: the repo distributes GPL-3.0 FFmpeg (with x264/x265) under a custom license — a GPL compatibility conflict documented in Issue #596 with no official response.


Key Metrics Summary

MetricValue
Stars15,585
Docker image size~70GB
Minimum GPUNVIDIA RTX 4070
Minimum RAM32GB
Minimum storage130GB+
Supported languages8
Last push2026-04-21
Commercial MAU threshold1,000 users

Verdict

Duix-Avatar packages mature open-source components (FunASR + Fish-Speech) with a proprietary C video engine — solid engineering execution, genuinely offline. But four hard problems:

  1. High hardware barrier: RTX 4070 + 32GB RAM + 70GB image — not a casual tool
  2. New GPUs unsupported: Blackwell (RTX 50 series) is broken, no maintenance activity for 5 months
  3. Non-open license: 1,000 MAU triggers commercial licensing; GPL conflict unresolved
  4. Local quality is one tier below cloud: Official admission — quality-conscious use cases still need to pay for cloud

Good fit: developers with idle NVIDIA GPU workstations wanting to experiment with local digital human technology. Not suitable: production deployment, mobile, or quality-critical use cases.


Open source for learning only. Check license terms carefully before commercial use.

💬 评论与讨论

使用 GitHub 账号登录后发表评论

关于本站 · 免责声明

🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18,不接受商业合作、不代表任何企业或机构立场,也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息,而是给出有独立思考的深入分析,希望帮更多人获得有价值的一手科技认知。

⚠️ 文中介绍的开源代码与模型,仅供学习交流与技术借鉴。它们大多仍处于早期阶段,有待进一步研究和验证,请勿直接用于工作或生产环境;如需采用,请先自行充分测试,并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.

  1. 本站文章均为作者基于公开信息的个人研究与观点整理,不代表文中提及的任何公司、产品、模型的官方立场,未与其构成商业关联或合作关系。
  2. 科技行业信息更新极快,我们尽力保证内容准确、及时,但不对完整性、实时性做绝对保证,具体请以相关企业/项目官方公告为准。
  3. 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有,我们会尽量注明来源;如你认为存在版权疑问或侵权,请通过下方邮箱联系我们,收到通知后会尽快核实处理(更正、加注来源或删除)。
  4. 文章内容仅为技术科普与个人观点,不构成投资、法律或其他专业建议,据此进行任何决策的后果需自行判断和承担。

📮 侵权 / 勘误 / 合作咨询:hello@mushroom.cv