HairDeck turns one character image into 3D hair cards with spring-bone chains, and its checkpoints are Apache-2.0. This week is 3D research, and HairDeck is the one piece a small studio can install today.
HairDeck turns one character image into game-ready 3D hair
Illustration generated with AI on our own machine.
What happened. Yuze He, Yixiang Wei, Wang Zhao, Chuhao Chen, Yanning Zhou and Yong-Jin Liu released HairDeck, which turns a single character image into stylized hair cards with per-card UVs, colours and spring-bone chains (Hugging Face). It exports GLB files and a VRM 0.x avatar, and the checkpoints are released under Apache-2.0. On CHARM’s own test set, the authors say it lowers Chamfer distance by 44% (HairDeck).
What we think. We think this is useful work, not a demo: hair is the slow, fiddly part of building an avatar. It runs on Linux and native Windows with Python 3.12, PyTorch 2.8 and CUDA 12.8, and Apple silicon support arrived on 3 November (GitHub). The downloader even sets up Blender 4.2 LTS with the VRM add-on.
The good bit. Apache-2.0 means a studio can use the output commercially.
What you can do. If you make avatars or game characters, try it on ten of your own drawings and count how many need fixing by hand. Check your graphics card first in our VRAM calculator.
Nvidia’s Lyra research builds 3D scenes from images and video
Illustration generated with AI on our own machine.
What happened. Nvidia researchers published LucidLyra, which distils a camera-controlled video diffusion model and frozen 2D vision models into 3D Gaussians (Nvidia). It also reconstructs structured, moving 3D scenes from a single video. Separately, heise reports that Nvidia released Lyra 2.0 as a research project on GitHub and Hugging Face, exporting Gaussian Splatting and mesh files that can go into its Isaac Sim robotics simulator (heise).
What we think. We think both are robot training grounds first and game tools second: Isaac Sim is where the scenes are meant to end up. heise points out that Google’s Project Genie is only available in the USA, while Lyra 2.0 is on GitHub (heise). Nvidia sells the GPUs, so giving the software away costs it nothing, and you get to keep the files.
The good bit. Mesh export means the output opens in ordinary 3D tools.
What you can do. If you build simulations or product visualisations, keep an eye on it. For running research code on your own box, start with our local AI for developers page.
FlowHMR recovers 3D human motion from a single video
Illustration generated with AI on our own machine.
What happened. The FlowHMR authors present a framework for recovering physically plausible global 3D human motion from monocular video (Hugging Face). On their Wild-4K experiments, FlowHMR reached a physical tracking success rate of 82.47%, against 62.82% for the GVHMR baseline.
What we think. We think motion capture from a phone video is a real saving for small animation and sports teams, if the 82.47% holds outside the authors’ own dataset. “Physically plausible” means the feet stay on the floor, which is the first thing viewers notice when it goes wrong.
The good bit. One ordinary camera instead of a suit and a studio.
What you can do. Nothing to install from this page yet. To see which machine suits this kind of work, read our hardware by job guide.
ReLift repairs 3D point tracks from video, faster than 4RC
Illustration generated with AI on our own machine.
What happened. ReLift treats dense 3D tracking as repairing and completing trajectories, using only geometric inputs (ReLift). On WorldTrack, the authors say it improves on the strongest baselines by 2.1 APD on all points and 3.0 on dynamic points. They also say its faster configurations run 2.7 to 4.4 times faster than 4RC.
What we think. We like the idea of fixing a tracker’s output instead of training a bigger tracker. The corrector is trained on 3D trajectories with simulated errors, without image-based training (Hugging Face), which keeps it small.
The good bit. Faster tracking is what makes measuring movement in video affordable.
What you can do. Nothing for an office yet. If you work on video analysis, our local AI for developers page covers running such tools privately.
VoxLution generates detailed 3D rooms
Illustration generated with AI on our own machine.
What happened. VoxLution is a coarse-to-fine transformer approach that generates detailed indoor scenes, refining scene chunks down to 1.5625 cm voxels (VoxLution). Its authors report 59.6% lower FPD and 51.1% lower FID on 3D-FRONT than existing baselines, and 70.5% lower FPD and 73.8% lower FID on ScanNet++.
What we think. We read those percentages as the authors’ own comparison, and generous ones. Generated rooms are useful for training robots and for showrooms, less so for an architect who needs real measurements.
The good bit. Centimetre-level voxels mean the furniture looks like furniture.
What you can do. Watch it if you sell interiors online. For the business side of Spanish AI adoption, see our look at the Spain AI market opportunity.
In one line
HairDeck is the 3D release worth installing this week: try it on ten of your own drawings before you trust it.
Work with us
If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.