A standing portrait grows swaying hair: how HairDeck gained 318 stars in a day from the Japanese VRM scene

HairDeck is paper code from a PhD group at Tsinghua. Give it one anime character image and it generates individual hair strands, then a bone rig that makes the hair sway, exportable as a VRM avatar. The repo is 8 days old and has 849 stars. The stars didn't come from Hacker News or big English AI accounts. They came from a Japanese 3DCG news account. It has many research peers, but it got stars because what it outputs is immediately usable by VTubers and VRChat players.

A Nov 1 noon post

2026 年 11 月 1 日中午 12 点半(日本时间),日本 3DCG 资讯站 MODELING HAPPY 发了一篇文章,X 上同步发帖,标题是「HairDeck——从一张图自动生成动漫风 3D 头发和让头发摆动的骨骼!支持导出 GLB、VRM,免费 demo 已公开」。帖子把卖点列成几个勾:骨骼、毛束形状和 UV、VRM、免费 demo。文章里专门提了一句:一般的单图生成 3D,头发会和身体糊成一整块网格;HairDeck 保留了一片片独立的发片,每片都带 UV。官网演示里,角色一转头,头发晚半拍才跟过来。

This post hit 23.8k views, 440 likes, and 463 bookmarks (on X, 11-06 to check). More bookmarks than likes means people are saving it to come back to. The same day, a Bilibili video called "HairDeck single-image stylized hair package" appeared: 1 min 50 sec, and it has 21.3k plays as of 11-06.

The next day, the repo went from 262 to 580 stars, +318 in one day, its biggest day so far.

What it actually does

For people making VRChat avatars or VTuber models, hair is always the hardest part. Anime hair in games and avatars isn't individual strands. It's dozens of textured "hair cards" layered together, and they need spring bones to sway. Until now there were two options: hand-paint each card in VRoid Studio (via pixiv), or buy ready-made hairstyle assets on BOOTH, which will never match the character you drew.

HairDeck turns that step into "upload one character portrait." The model treats the whole hairstyle as a set of hair strands, generates them all at once, then uses a second model to generate the bone chain that drives them. Finally it calls Blender and a VRM plugin to attach the hair to the avatar and export. Starting Nov 5, it can also read colors from the image, so two-tone hair and hair accessories come through.

The author, Yuze He (on GitHub as hyz317), is on his third anime-character project. The earlier ones were StdGEN, which generates a whole character from one image (406 stars), and CHARM, anime hairstyle generation (60 stars). The paper says HairDeck's geometric error is 44% lower than CHARM's and it runs 33 times faster, but almost no one sharing the post mentions those numbers.

How the stars grew

Data: GitHub stars tracked daily; on 11-06 the GitHub page showed 849 stars and 60 forks.

A few things to note:

The Japanese community picked it up first. The author had a Japanese README from day one, plus Chinese README and download notes for local weights. Japanese sites and Bilibili got more traction than the English side. English coverage was just @Stefan_3D_AI, an aiweekly news item, and two small YouTube videos. Hacker News had nothing, and Reddit only one reposted link.

Updates never stopped. The online demo worked on day one. After that came 11-02 for Windows, 11-03 for Mac, and 11-05 for color, roughly one update every day or two, each giving newsletters and package makers a reason to post again.

The spike was one-time. After the peak, the repo gained only 20 to 40 stars a day. It's the typical curve of something carried by one event, then slow growth.

Who's doing something similar

"One image → stylized hair cards + bones + VRM" hasn't shown up from a second player in either academic projects or commercial products. The followers so far are wrappers, not clones: one ComfyUI node (2 stars), one Pinokio one-click installer, one Bilibili package.

The clearest signal is CHARM vs. HairDeck: same author, same topic, and not a big research gap. One has 60 stars, the other 849. The difference is whether the output drops straight into tools players already use.

What's still in the way

Running it locally needs an NVIDIA GPU (or an Apple Silicon Mac), and you first have to request Meta's gated SAM 3 and DINOv3 weights on Hugging Face. Normal players will struggle to set it up. One YouTube video is titled "I Tested the Space (and It Broke)." The repo has two issues total: one person reporting a finished ComfyUI node, and another asking whether the mesh can be converted into Unreal Engine groom hair. How many people actually put the generated hair into their own avatars isn't visible from outside.

If it were you

My read: The lesson here is "ship the output in a format players can use directly," not "build a better hair model." Copying HairDeck directly won't earn stars. Using its approach for the next VRM part is a real opportunity.

Who it's for: Grad students and research groups working on 3D generation and character generation who have the models but lack distribution; or developers who know the Blender, Unity, and VRoid pipelines and can write plugins.

Who it doesn't fit: People who can't train models and only want to make a wrapper. The ComfyUI node with 2 stars already shows a bare wrapper won't get much. It's also not for anyone trying to beat it head-on in hair. The original author has already published three papers on this, and the dataset is his too.

How to do it: Three paths, ordered from most to least confidence:

1. Target the same player ecosystem with other parts: generate clothing, accessories, tails, or animal ears from one image as VRM items, with bones and direct VRM export, following HairDeck's format. (My inference: nobody is doing this yet. The hard part is training data.)
2. Fill the gap HairDeck left: issues show people asking about Unreal groom, and Unity, Unreal, and VRoid import plugins are still empty. Or build a version that works after install without requiring gated weights (check the SAM 3 and DINOv3 licenses first).
3. If you already have a paper stuck at the research-representation stage, add the step "export to a file a specific tool can open" before you publish.

How to get users: Ship code, weights and a playable demo on day one; write the README in Chinese, Japanese and English; make the demo a short video where the effect is obvious at a glance (like hair swaying when you turn your head); hand it directly to Japanese 3DCG news sites like MODELING HAPPY and shader.jp, and to Bilibili uploaders who make bundles; then post a platform or feature update every day or two. HN and English Reddit didn't do much for this kind of thing. HairDeck didn't rely on them.

Before when: If you're building downstream plugins, move now while HairDeck's buzz is still hot, roughly the next two to three weeks. Other VRM part ideas aren't as urgent, but once the big players' image-to-3D tools start splitting parts and exporting rigs, this gap closes.

What would mean I read it wrong: (1) A month from now HairDeck still gains only a dozen stars a day, and nobody on BOOTH or X is posting avatars made with it. That would mean players bookmark it without really using it, and "matching the player format" is overrated. (2) Someone builds a Unity or Unreal plugin and it's still under 100 stars after a month. That would mean downstream wrappers don't earn stars. (3) Meshy, Tripo, or Hunyuan3D ships "with bones, split into hair cards." That would mean this category gets absorbed by big companies quickly.

Data and sources: GitHub repo page, README change log, issues, and daily star history (11-01 to 11-06); arXiv 2610.40599 and project page; Hugging Face model, Space, and Daily Papers pages (11-06); MODELING HAPPY 2026-11-01 article and X post @happy_modeling; X posts from @Stefan_3D_AI (10-31) and @AICUai; Bilibili BV1ABxs6yEMa; comfyui-wiki, shader.jp, modelingtechnote, and aiweekly coverage; visualbruno/ComfyUI-HairDeck and the Pinokio page; competitor GitHub star counts (11-06).