Hair cards are one of the least glamorous parts of character art. They’re the flat, ribbon-shaped polygon strips that stand in for strands of hair in almost every real-time game or avatar system, because simulating thousands of individual hair fibers at 60 frames per second isn’t realistic. Building a good set of cards — placed, shaped, and rigged with the spring bones that make hair sway and settle — is skilled, time-consuming artist work. A new paper, posted to arXiv on October 29 by researchers including Yuze He and Yixiang Wei, proposes HairDeck: a model that generates that entire package from a single image.
What it outputs
Feed HairDeck one image of a stylized character, and it returns two things together: the hair cards themselves (the ribbon geometry) and the spring-bone chains that drive how those cards move. That second part matters — most prior image-to-hair methods stop at a static mesh or a pile of strands, leaving rigging as a separate, manual step before the hair can animate at all. HairDeck’s output is meant to plug directly into a game or animation pipeline without that extra pass.
The actual technical move
The one prior method built specifically for this — called CHARM — generates hair cards autoregressively, as one long sequence of control points, card after card, in order. HairDeck’s authors make a simple but effective reframing: a card hairstyle isn’t really a sequence at all. It’s a set of independently placed pieces an artist would lay down without a fixed order. So HairDeck represents each card as a single token and uses a set diffusion transformer to denoise — generate — all the cards together, in one pass, rather than building them one at a time.
That change in representation is also what drives the speed gain: generating a set jointly sidesteps the sequential bottleneck of autoregressive generation, where each new card has to wait on every card before it.
The numbers
Measured against CHARM on CHARM’s own test set, HairDeck:
- Reduces Chamfer distance (a standard measure of geometric accuracy) by 44%
- More than halves card-orientation error
- Runs 33 times faster
It also outperforms image-to-3D pipelines and strand-based hair methods more broadly when tested on single images, not just against CHARM specifically.
Why this matters beyond one benchmark
The bottleneck in stylized 3D character pipelines is rarely the rendering — it’s the labor of building assets that look right and animate correctly, over and over, for every new character. A method that collapses “photo of a hairstyle” into “riggable, game-ready hair asset” in one step, at a fraction of the compute cost of the prior approach, is the kind of result that tends to show up in production tooling (Unity and Unreal plugins, specifically) well before it shows up in a research demo reel. Worth watching for anyone doing character work in Houdini, Blender, or a game engine.
Related Reading
- HairDeck: Animation-Ready Stylized Hair Cards from a Single Image — arXiv:2610.40599
- AMD Cut the Memory Cost of Ray-Traced Foliage by 47x. Then an AI Wrote a Path Tracer in an Hour. — our coverage of another graphics-pipeline efficiency result
- Blender Python: Getting Started With bpy Scripting — our tutorial, relevant to anyone automating character asset pipelines
- Marvelous Designer: Getting Started With Cloth Simulation — our tutorial on a related stylized-character asset workflow