Skip to content
IFAI

Open source

HairDeck turns a single character image into animation-ready 3D hair

Published Reach 44

Researchers just open-sourced HairDeck, a model that turns a single 2D character image into stylized 3D hair cards complete with spring bones. It generates physics-ready hairstyles in about one second—33 times faster than previous methods—and exports directly as a VRM avatar. At just 2.45GB, it lets creators generate animatable hair on consumer GPUs.

HairDeck turns a single character image into animation-ready 3D hair
Image: hyzcluster.github.io

If you have ever tried turning a 2D character into a 3D avatar, you know the hair is always a nightmare. Most AI 3D generators spit out a stiff, fused blob of mesh that looks like a plastic helmet and refuses to move when your character turns around.

Traditional game studios solve this using hair cards—flat, textured ribbons placed by hand, rigged with spring bones so they bounce naturally. Now, researchers have open-sourced HairDeck to automate that entire workflow.

You feed it a single 2D character illustration, and it outputs a complete, layered set of stylized 3D hair cards with UVs and physics bones already attached.

One card, one token, 33 times faster

Until now, the leading method for generating stylized hair cards was a system called CHARM. CHARM treated hair as a sequential string of control points generated autoregressively, which made it painfully slow and prone to weird orientation errors.

HairDeck approaches the problem differently. A hairstyle is really an unordered collection of ribbons. The researchers represent each hair card as a single token covering its shape, root position, and orientation, then use a set diffusion transformer to denoise all cards together in one shot.

A second transformer then builds the spring-bone chains, binding the cards through geometric skinning. The pipeline even samples colors straight from your input image, faithfully capturing highlights and multi-colored anime hair.

The benchmark numbers on CHARM's own 100-hairstyle test set are wild: card generation time dropped from 33.7 seconds to 1.01 seconds—a 33x speedup. Chamfer distance dropped by 44%, and card-orientation errors were cut by more than half.

2.45GB, runs locally, exports straight to VRM

What makes HairDeck particularly exciting is how accessible the researchers made it. This is not a locked paper demo; the code, weights, and training scripts are already live on GitHub and Hugging Face.

The model weighs just 2.45GB. It runs on consumer Nvidia GPUs and even supports Apple Silicon out of the box, meaning you can easily spin it up on a MacBook.

For VTubers, indie game devs, and 3D animators, the generated hair exports directly as a VRM avatar. When the character moves or turns its head, the hair sways automatically with dynamic bone physics.

To top it off, the project is licensed under Apache 2.0 for unrestricted commercial use, and the team provided full training scripts so studios can fine-tune it on their own art styles.

Why it matters

Most generative 3D tools obsess over outputting a single solid object, but those meshes are almost useless in production pipelines because they are static and messy. HairDeck is smart because it focuses on a modular, production-standard primitive: stylized hair cards with dynamic bones.

For indie creators, this takes a task that usually takes hours of manual card placement and bone painting in Blender and shrinks it down to a few seconds. The output is genuinely animation-ready rather than just a decorative render.

The main caveat is style coverage. HairDeck is tuned specifically for stylized and anime-like hair ribbons; you will not get realistic strand physics or intricate braids out of it just yet. Still, baking physics bones directly into the diffusion process is a brilliant move that sets a new baseline for practical character generation.

Sources (1)

  1. [1]
    hyzcluster.github.io(opens in a new tab)Originalhyzcluster.github.io

Best of AI, daily