BaoCut Explains Virtual Streamers Like Shadow Puppets — 3 Steps to Add One to Your Videos

dotey · x · 2026-11-11

A breakdown of how the BaoCut virtual streamer works, explained as shadow puppetry: AI draws a front-facing portrait, modifies it 11 times with pixel-level diffing to extract 27 layers; a mesh over the whole image enables head turns, with 17 spring-like knobs for hair; mouth shapes are chosen per-character pronunciation, opening 0.03s before audio. Rigging is ported from open-source LivLiv Player (Apache-2.0); the avatar was generated with Seedream 5.5.

Setup in 3 steps

Run baocut skill install so agents like Claude Code/Codex can drive it; the built-in AI assistant also ships the skill
Pick from 19 bundled cartoon avatars, or have the agent draw a front-facing portrait (eyes open, mouth closed, plain background), then modify/layer/package it — 9 image generations
Give it a topic; the agent writes the script, voices it, computes lip-sync, and returns a review link; re-dubbing only re-computes mouth shapes

Limits: cartoon characters only, no real-person lip-sync or face swap; best lip-sync for Chinese/English/Japanese/Korean; max 30° head turn, no hands, no bald characters.

Quoted context: ByteDance's Nov 9 Seedream 5.5 keeps untouched regions pixel-identical across 100 consecutive edits, generates transparent-background images, cut severe text error rates from 2.8% to 0.4%, and beats rivals on text accuracy vs GPT Image 2.5 Sunburst and Nano Banana 2.1. 2K square edits cost 0.28 yuan via Volcano Ark.

Original post →

BaoCut Explains Virtual Streamers Like Shadow Puppets — 3 Steps to Add One to Your Videos

More from Multimodal