The Digital Human Tutorial He Spent Thousands of Dollars Building Is Now Open Source

Berryxia posted on X: the upgraded version of the digital human tutorial is here. The code is already on GitHub under the MIT license, open source and free. He says he worked tirelessly for weeks and burned through thousands of dollars, with the goal of letting beginners produce consistently.

The project is called digital-human-koubo. Its core capability: one photo plus one sentence automatically generates a complete digital-human talking-head video.

This workflow covers scriptwriting, lip sync, voiceover, subtitles, and editing/packaging. Your Agent first diagnoses the content, recommends 2 of 5 editing styles, asks one round of questions, then writes the script, breaks it into segments, quotes the cost, generates, runs quality checks, unifies the voice, and finally renders the video with Remotion.

The entire generation pipeline is completed through Renoise MCP, the interface between the Agent and the video generation model, calling Grok Video 1.5. Each video segment is no longer than 15 seconds; costs are quoted before generation, and batch runs happen only after confirmation. After generation, Whisper checks each segment against the script and also performs frame sampling checks. Grok occasionally burns subtitles into the frame, and such clips are regenerated.

On cost, the author provides real test data. A 27-second live-commerce sales ad, with 1 scene image plus 5 Grok clips, cost 482 credits in total. A 4-minute 36-second tutorial video with 32 clips costs about 3,250 credits. At Grok Video 1.5's current rate of 17 credits per second at 720p, one minute of talking-head video costs just over 1,000 credits. Credits are the compute quota on the Renoise platform.

The part of this project that took the most thought was packaging. The author analyzed 25 top creator videos frame by frame, including MrBeast, MKBHD, Media Storm, He Tongxue, Johnny Harris, and Banfo Xianren, and extracted 48 reusable packaging components in four groups: opening hooks, data presentation, annotations, and pacing/engagement. For example, TextBehind can place large numbers between the person and the background, using macOS Vision for portrait segmentation and three-layer compositing. Components are invoked as JSON pack lists and automatically adapt to 9:16 vertical screens.

The 5 editing styles cover different scenarios:

• A: finance/tech explainers, suited to "why" content
• B: influencer talking-head videos, vertical, for product sales, recommendations, and opinions
• C: new product breakdowns, 30-90 seconds
• D: step-by-step tutorials, 3-8 minutes
• E: multi-host, multilingual, multi-scene

Six languages were tested: English, Japanese, Korean, French, Spanish, and Arabic, with Whisper verification accuracy of 94-100%. For scenes, GPT Image is first used to composite the digital human into any background—a live-commerce studio, office, classroom, or Kyoto street—and then the digital human is made to speak. The male digital human in the demo video is the author's own digital twin, made from his own photo.

This workflow is released as an Agent Skill; essentially, it is a set of instructions plus scripts that, once installed into an Agent, teach the Agent the entire process. Installation is not complicated: clone the repo, put the two skill folders into the Agent's skills directory, and connect to Renoise MCP. It has been tested and works in Claude Code; Codex should theoretically work with the same scripts. WorkBuddy, Doubao, Qwen Office, and Nano Work also support it, but it depends on whether the product fully supports three capabilities: loading Skills, connecting to remote MCP, and running local commands. If the Agent can only run in a cloud sandbox, generating clips is fine, but the local editing step requires Claude Code or manually running commands.

Some users reported after testing that the voice and emotion still sound somewhat robotic. The author also acknowledges this in the documentation. To use your own voice, you can connect Xiaomi MiMo TTS for voice cloning. However, Grok generates lip sync based on the native voice; switching voices causes a lip-sync offset of about 100 ms. It is not obvious in medium shots but can be seen in close-ups. For perfect lip sync, you need audio-driven models such as OmniHuman.

The author does not claim there is zero barrier. He offers this judgment: in digital-human solutions, the tension between efficiency and quality has always existed, and the point of open-sourcing is to give the choice back to users. You can use your own Agent, your own account, and your own credits to run your own pipeline.

Repository: github.com/andyhuo520/digital-human-koubo
Renoise MCP installation: renoise.ai/mcp