How BaoCut turns 2D images into puppet-style AI avatars

Most people trying to build an AI talking head today reach for heavy video diffusion models. The result is predictable: GPUs melt, costs soar, and if you change a single word in the voiceover, you have to re-render the whole clip from scratch.

BaoCut is taking an entirely different approach. The tool recently broke down the mechanics of its virtual avatars, and in plain terms, it isn't generating video at all. It is staging a digital puppet show.

27 layers and spring physics

The whole system breaks down into three simple steps: slice, move, and speak.

First comes the slicing. An AI model generates a single front-facing portrait, followed by eleven subtle variations. The program compares the pixels across these versions to automatically separate the character into twenty-seven distinct layers. In its demo, BaoCut used ByteDance's Seedream 5.5 to generate the base art.

Next is the motion. BaoCut ported the skeletal deformation and spring physics from LivLiv Player, an open-source project licensed under Apache-2.0. The character is draped over a 2D mesh controlled by seventeen parameter knobs, allowing the head to turn and giving the hair bouncy, spring-like secondary motion.

Finally, there is speech. Instead of watching the audio volume to flap the mouth open and closed, the engine reads the exact pronunciation of each syllable to pick the correct phoneme. It even triggers the mouth shape 0.03 seconds before the audio plays, matching human speech physiology.

Because the output is driven by 2D layer transformation rather than frame-by-frame pixel generation, it runs without heavy graphics cards. If a script changes, the engine simply recalculates the lip shapes without redrawing the character.

Turn coding agents into video editors

Beyond the rendering trick, BaoCut packaged this entire pipeline into an automated skill. Running baocut skill install in a terminal connects the capability directly to agents like Claude Code or Codex.

Users do not have to mess with keyframes or timelines. You can simply pick from 19 built-in original cartoon avatars, or have the agent generate a custom character from scratch.

Building a custom avatar takes about 9 image generation cycles. The prompt requires a front-facing face with eyes open, mouth closed, a solid background, and hair. Once you approve the base image, the agent edits, cuts the layers, and packages the avatar bundle.

From there, a single prompt can trigger an entire production. You can tell the agent to build a three-minute explainer with an avatar in the lower-right corner. It writes the script, records the voiceover, computes the phonemes, drops the assets onto the timeline, and gives you a preview link.

There are clear boundaries, though. BaoCut strictly handles cartoon characters—no photorealistic faces, no lip-syncing over real humans, and no face swapping. Pronunciation mapping is precise for Chinese, English, Japanese kana, and Korean, while other languages fall back to basic volume-based mouth opening. Characters cannot turn their heads past 30 degrees, have no hands, and bald avatars are currently not supported.

Why it matters

For solo creators and developers making product demos, this drops the production barrier right to the floor. Traditional AI talking-head videos force a brutal tradeoff: burn significant money on heavy neural video rendering, or settle for boring slides. BaoCut shows that clever 2D layer engineering can deliver lively presentation without touching heavy video models.

We think BaoCut's smartest move was decoupling asset creation from real-time motion. Instead of fighting for scarce GPU compute to generate video frames, it leverages lightweight mesh deformation and precise phoneme matching. Exposing this via an Agent skill is particularly clever: it turns coding agents into fully automated editing suites that can draft, voice, and cut a video without human micromanagement.

The obvious limitation is the ceiling of 2D puppets. With a strict 30-degree rotation limit, no hands, and cartoon-only styles, this will not replace high-end video production. But for software tutorials, news digests, and corner explainers where low latency and zero rendering fees matter most, this is easily one of the most practical setups available right now.