talkframelisted
Install: claude install-skill sf-ryanyan/talkframe
# TalkFrame — photo + voice → talking digital human
Make a lip-synced talking-head video from **one portrait photo** and **a script or an audio file**. Works with cloud APIs (fal.ai, ListenHub) or a local GPU. Built for Chinese users: **all user-facing messages are in 简体中文**; prompts to models are English; file names are English.
Skill folder = `<skill>` below. Scripts: `<skill>/scripts/`. Provider details: `<skill>/providers/` — read **only** the file for the route you use (saves tokens).
Short names in commands: `tf_media.py` = `python <skill>/scripts/tf_media.py` (same for `tf_fal.py`, `tf_check.py`; use `python3` on Mac/Linux). Every script prints one JSON object; read it instead of guessing.
## What "good" means (in priority order)
1. **口型对** — mouth shapes match the words and timing; mouth closes in pauses.
2. **脸不变形** — face shape, eyes, nose, teeth, hairline stay like the photo.
3. **动作克制** — blinking, breathing, at most a slight head movement. No big smiles, gestures or body turns.
4. **不浪费** — check inputs first, show cost/time, get a yes, test short before long, never auto-retry a paid task.
## Two modes
| Mode | How | When |
|---|---|---|
| **自然模式 natural** (default) | one-step model: photo + audio → video | good front-facing photo |
| **保守模式 conservative** | still/near-still base video → **mouth-only** lip sync | risky photo, OR natural result failed QC. "不笑、不动,只张嘴说话" |
Auto fallback: a risky photo starts in conservative mode; a natural result that fails Q