Juicebro-star · Budget Digital Human Talking-Head
Repository: https://github.com/Juicebro-star/guozhige-koubo ｜ Each finished video costs about 1.36 USD
One Sentence for the User
"Help me run this repo and produce a 15-second talking-head video with the topic ___"
What you need to do: Fill in pipeline/config.json → Run 1_prompt.py → Ask the user to generate in the H3 console (quote 1.35 USD first) → Run 2_make.sh → Deliver mp4 + publication copy.
When to Use
The user wants a vertical (9:16) digital human talking-head video: a "person" speaking to the camera for 15 seconds, with stylized text, keyword color changes, and a QR code.
Process (Five Steps)
- Edit Script + Change Avatar → Edit
pipeline/config.json(script/avatar/look/title/keys/big/stack/pill) - Generate Prompt →
python3 pipeline/1_prompt.py - Generate → User generates in the Meta H3 console using the prompt (9:16 / 768P / 15 seconds, approx. 1.35 USD), download the mp4
- Produce Video →
bash pipeline/2_make.sh <final_video.mp4> - Publication Copy → Provide according to platform (Douyin/WeChat Channels/X have different logic)
See pipeline/README.md for details.
Ready-Made Script Templates
Four templates in pipeline/templates/ — just tweak a couple of sentences:
| File | Scenario |
|---|---|
01-tool-review.json | Review a tool |
02-pain-point-sharing.json | Share experience/lessons |
03-group-Q&A.json | Answer frequent questions |
04-single-tip.json | Cover just one point |
CONFIG=templates/01-tool-review.json bash 2_make.sh ~/h3.mp4
Change Style (No Re-Render, Zero Cost)
The same footage can output different stylized text styles. pipeline/variants/ contains ready-to-use configurations:
cd pipeline
CONFIG=variants/A.json bash 2_make.sh ~/Desktop/h3.mp4 # Subtle, realistic style
CONFIG=variants/B.json bash 2_make.sh ~/Desktop/h3.mp4 # List-based, practical style
This Skill Is Self-Contained
SKILL.md + persona.json + assets/ + scripts/ + references/ + templates/ + pipeline/ — just copy the entire directory to use (approx. 10MB, excluding runtime dependencies).
The first run of 2_make.sh will install rendering dependencies in pipeline/work/ (about 1 minute, reused afterward).
Hard Rules
- Single segment cannot exceed 15 seconds: H3 bills double for anything over 16 seconds. Script should be 60–67 characters.
- Script must not exceed 67 characters: 4.5 characters/second. If exceeded, either cut or rewrite.
- Do not fabricate numbers or experiences: Never write "Someone earned X monthly with this" for cases that don't exist.
- Must label AI: Add corner watermark + platform declaration (mandatory on WeChat Channels).
- Only use authorized faces and voices: No celebrities, no unauthorized assets from others.
- QR code must be black: Orange QR codes have insufficient contrast and may be unscannable after compression.
Script Writing (Lessons Learned)
- Avoid negative structures like "It's not A, it's B" — instantly screams AI-generated.
- Start with the result or scenario, not preamble: "People in the group discuss every day: how to actually use AI."
- Mid-section: give three concrete items, with specific nouns: "Tool selection, Agent setup, job monetization."
- End with one action, without pressure: "Want to take a look? Scan the code to join."
Stylized Text Configuration (config.json)
| Field | Purpose |
|---|---|
big | Large orange text in the middle section, fill at with the word from the script |
stack | Three items popping up one by one on the left, at is the word for each |
pill | Capsule above the QR code (scan to join group) |
keys | Words to color-change in subtitles, must appear verbatim in the script |
qr_pos | Position of QR code in 1080×1920 frame |
preset | Stylized text intensity: subtle (realistic sharing vibe) / standard / lively |
Pre-Delivery Checklist
- Check 4 frames: verify no burned subtitles, no face drift, correct lip sync
- Scan the QR code on the finished video (not the original image) with another phone
- Ensure video length is ≤15 seconds