Hair Physics in Gaming, 2026 Update: Simulation, Neural Rendering and Coiled Hair
Why an update
In July 2025 I published Hair Physics in Gaming: Why Realism Remains an Elusive Quest. This article brings it up to date with developments between July and September 2026, and corrects an error.
That post credited NVIDIA with a neural hair model called "NeuralWigs." I can find no NVIDIA work by that name. The closest match is NeuWigs, a 2023 research model from a team working at Meta, built to capture and animate hair for virtual-reality avatars rather than for games. I have also corrected the history of strand-based hair in game engines. Where a statement below rests on my own observation rather than published evidence, I say so.
Though I am outside the gaming industry, I am fascinated by how my industry intersects with the visual aspects of the gaming experience on multiple levels. I have spent 39 years in the hair and makeup industry watching hair move under lights, on stage and on camera, in photographs, under almost every condition imaginable. What interests me about games is that they are attempting, in roughly 16 milliseconds per frame, something my profession has had to understand: that hair responds to its environment, to motion and to a myriad of variables, even at rest.
I welcome corrections and advice from the artists, engineers and physicists who work in games.
Below is my own reading of what is currently possible, and of the problems that remain.
The physics problem: what hair actually does
A healthy human scalp carries on the order of 80,000 to 120,000 terminal hairs (StatPearls, Anatomy, Hair). Game characters are now built on a comparable scale. Epic Games describes its strand system as able to render and simulate hundreds of thousands of individual hairs in real time, and NVIDIA reported that Geralt in The Witcher 3 (2015) carried an average of 30,000 hairs, rising to 115,000 in close-ups, with around 6,000 in his beard alone.
Each strand is a thin, flexible fibre with an internal structure: a protein cortex wrapped in overlapping cuticle scales. It bends and twists easily, stretches, collides or entangles with its neighbours and with the body, and responds to gravity, wind, water and movement. Simulating that for every strand, every frame, is a large coupled-dynamics problem.
The standard answer in games is guide-strand simulation. The engine simulates a smaller set of guide strands and interpolates the rest from them; Unreal Engine exposes this directly as a guide-density setting and interpolation controls. It is an effective economy. Its cost, in my observation as a viewer rather than as a measured finding, is that hair often looks convincing at rest and less convincing in motion, precisely where a wig maker or hairstylist would look first.
At 60 frames per second, every frame must be produced in about 16.7 milliseconds (1,000 ms ÷ 60); at 120 fps, about 8.3. Hair competes for that time with lighting, shadows, animation and everything else on screen. I have seen figures quoted for the share of the frame that hair consumes, but none I could trace to a published measurement, so I have left them out.
Why curls, coils and textured hair are harder
Straight hair is expensive to simulate but comparatively predictable. Curly, coily and afro-textured hair adds difficulty at every stage, and the graphics research literature now says so explicitly.
Geometry. A tightly coiled strand's resting shape is already a helix, often with reversals in direction. The Yale Computer Graphics group has devoted a series of papers to this alone — Lifted Curls (2023), Curly-Cue (2024) and, in July 2026, a paper on editing coiled hair in terms of curvature and twist (discussed below).
Volume and contact. Coiled hair forms a mass in which strands support and push against one another even at rest. That is exactly the interaction guide-strand interpolation tends to smooth away.
Cheaper representations break down. As I understand it, most games still rely on hair cards, textured strips standing in for groups of strands, for gameplay. Recent research on generating cards automatically reports that the approach becomes visibly polygonal for highly curled styles, and that coily hair loses fidelity when the number of cards is limited (Strands2Cards, 2025; an automatic card-extraction method, 2025).
Light. Coiled hair occludes and scatters light differently from a smooth, aligned sheet of straight hair. How best to model that in real time remains, as far as I can find, an open research question rather than a solved one. Moisture, natural oils, cuticle condition and styling products also change how hair takes the light, something every stylist adjusts for on set.
The outcome, for years, has been that characters with natural textured hair were harder to render convincingly than characters with straight or wavy hair. That is a technical problem and a representation problem at once. I return to the evidence on the second point below.
Three problems, not one
My July 2025 post treated "hair realism" as a single problem. Reading the 2026 work, I think it is clearer to separate three, because progress on each is happening at a different pace and in different places.
Simulation - how hair moves: gravity, inertia, collision, strand against strand.
Appearance - how hair looks: how light passes through, reflects from and scatters within a mass of fibres.
Modelling - how a hairstyle's shape is authored in the first place, before anything moves or is lit.
A wig maker or hairstylist would recognise the distinction. How a piece moves, how it takes the light and how it is dressed are separate judgements, and a failure in any one of them is visible. The most significant developments between July and September 2026 fall mainly under appearance and modelling. Simulation, arguably the part players notice most in motion, has advanced in research but not, as far as I can find, in shipped games.
Where strand-based hair stands in game engines
Strand-based hair in games has a longer history than my July 2025 post suggested. AMD's TressFX made its debut in Tomb Raider (2013), which AMD described at launch as the first in-game implementation of a real-time, per-strand hair physics system. NVIDIA's HairWorks was used in The Witcher 3 (2015). EA presented strand-based hair rendering in its Frostbite engine at SIGGRAPH in 2019. Unity's Enemies demo and, more recently, Indiana Jones and the Great Circle have both been the subject of technical presentations on strand-based hair.
Epic Games' Groom system in Unreal Engine predates Unreal Engine 5. It arrived as an experimental feature in Unreal Engine 4.24 in December 2019 and was declared production-ready in 4.26. It imports strand data in the Alembic format, renders individual strands, and drives their physics through Epic's Niagara system, falling back to cards and simpler meshes at distance. Epic's own documentation is notably modest about it, describing it as a minimal viable implementation of strand-based importing, editing, rendering and simulation, and noting that it does not provide tools for shaping the hair itself.
That modesty seems to me informative: it suggests where the industry actually is: strand hair is available and increasingly used, but most gameplay hair is still cards, and the tools for authoring convincing hair, especially coiled hair, are still being invented.
September 2026 - appearance: NVIDIA DLSS 5 ships
The largest industry event since my July 2025 post is NVIDIA's DLSS 5, released on 1 September 2026 in NBA 2K27 for GeForce RTX 50 Series graphics cards. I will be precise about what it appears to do, because it could be misread.
DLSS 5 does not change how hair moves. It is a neural network that runs as the final stage of rendering: it takes the frame the game engine has already drawn and adds lighting and material detail on top of it. Among the effects NVIDIA names are light transmission through hair and the way hair glows when lit from behind. In its announcement, NVIDIA also acknowledged the underlying constraint this post is about: in real time, individual hair strands are commonly replaced with simplified geometry.
Three cautions seem to me essential:
It is appearance, not physics. The strands beneath are simulated, or not, exactly as before.
It is costly. NVIDIA reports very high frame rates with its full suite of techniques enabled, including AI frame generation. Testing by Digital Foundry, reported by IEEE Spectrum, recorded performance drops of 40 to 60 percent. It currently runs only on RTX 50 Series hardware.
It is contested. Some players object when the model alters the look of characters beyond what the game's own renderer produced. NVIDIA emphasises developer controls, such as strength settings and masks, intended to limit exactly that.
One question I have not seen addressed anywhere: DLSS 5 made its debut in a basketball game whose players include many athletes with tightly coiled hair, locs and braids. How well a generatively trained appearance model renders those textures is, for now, an open question. I would like to see it evaluated.
July 2026 - modelling: coiled hair described by its curvature
The development I find most interesting this year came from research rather than from industry. At SIGGRAPH 2026 in Los Angeles (19–23 July), Alvin Shi and Theodore Kim of Yale, Florence Bertails-Descoubes of Inria in Grenoble and A.M. Darke of UC Santa Cruz presented Curvature Space Editing of Highly-Coiled Hair.
Their starting point is that tightly coiled hair is hard to shape with ordinary tools, which move points around in space. Instead, they describe each strand by its curvature and twist, building on super-helices, a model of hair as a chain of helical segments introduced in 2006 by Florence Bertails and colleagues. That earlier work was itself a collaboration between INRIA and L'Oréal Research and Innovation. Working in that space, an artist can expand, shrink, "ruffle" or interpolate a coiled style while it keeps behaving like coiled hair. Press coverage noted that the team drew on the cosmetics field, and that most earlier advances in animated hair served the straight or wavy hair most common to white-skinned characters.
The same conference included a production talk, Practical Methods for Kinky-Coily Hair in Animation Production, by Guillaume Pernin, Michael Davison and Curtis Andrus of Netflix Animation Studios.
Two honest qualifications. This is modelling and editing for animation, not real-time game physics. And the distance between a SIGGRAPH paper and a shipped game is, as I understand it, often measured in years.
What struck me, though, is the vocabulary. Describing a curl by measured curvature and twist, rather than by a type label, is close to how I think real fibres should be described. My own work on a measurement-based characterisation of sourced hair records one of several aspects of curl as a ratio of stretched to natural length. I make no claim of a connection between the two, only that the graphics field and the material field appear to be arriving, separately, at a similar conclusion: curl is better described by continuous measurement than by category.
2026 - simulation: neural dynamics, still in the laboratory
The idea that drew me to this subject in 2025 remains: instead of calculating hair physics from first principles every frame, train a neural network on simulated motion and let it predict what the hair should do, at a fraction of the cost. The research has moved on since then, but more slowly than my July 2025 post implied.
NeuWigs (Wang and colleagues, CVPR 2023), the model my July 2025 post misattributed to NVIDIA, learned to capture and animate real hair from multi-view video for virtual-reality avatars. Later work pursued real-time neural simulation more directly. Quaffure (preprint, 2024) produced real-time results but only for quasi-static hair — hair settling into a position, not swinging through motion.
Neuralocks (Lin, Larionov, Chen, Roble and Stuyck), published in Computer Graphics Forum in March 2026, addresses that limitation. Its authors describe earlier neural methods as limited to quasi-static solutions and report what they call the first successful use of self-supervised training for realistic, efficient dynamic hair simulation.
That is genuine progress. But I have found no shipped consumer game that uses neural hair simulation as of September 2026, though I may have missed one. The prediction I made in 2025, that learned hair physics would reach a major title within a few years, remains a prediction.
Representation: what the evidence now shows
My July 2025 post made claims about studio practice and industry advocacy that I could not trace to published sources, so I have replaced it with evidence that can be checked.
In 2022, Theodore Kim and colleagues at Yale and McGill argued, in a SIGGRAPH talk titled Countering Racial Bias in Computer Graphics Research, that the field's algorithms for virtual humans had targeted pale skin and straight hair. In June 2026, Kim, Alexa Schor, Julian Posada and Alka V. Menon published a systematic study at the ACM Conference on Fairness, Accountability, and Transparency (FAccT). They reviewed the 6,864 papers published at SIGGRAPH and in ACM Transactions on Graphics, the field's leading conference and journal, and report that the review confirms those earlier hypotheses about how humans are depicted.
The coiled-hair papers described above are, in part, the research response. They also suggest that the gap is not closed by hardware alone: the researchers found they needed methods designed for coiled geometry, not simply more computing power applied to methods built for straight hair.
Why does this matter to someone who sources physical hair? Because the properties that make coiled hair difficult to simulate (its curvature, its volume, the way strands interact with one another) are the same properties that determine how it behaves in a wig, in a prosthesis, on stage or on film. I have long thought that the digital and the material problems are the same problem viewed from different perspectives.
Where this may be headed
In 2025 I listed four expectations. Here is where each stands, as honestly as I can judge it.
Expectation (2025) | Status, September 2026 |
Neural simulation reaching a major game | Not yet, as far as I can find. Research advanced in 2026 (Neuralocks). |
Strand-level hair on console hardware | Not assessed here; I have not found a reliable source on current console titles. |
AI-assisted tools for diverse hair types | Partly. Research tools for coiled hair are appearing (SIGGRAPH 2026); production adoption is unclear. |
Game research and real hair science informing each other | Early signs: the coiled-hair research drew on the cosmetics field. |
One thing I did not anticipate: the first AI technique to reach players at scale acts on how hair looks rather than how it moves. Whether neural appearance and neural motion eventually meet in one system is, I think, the question to watch.
The quest for convincing hair in games is not purely aesthetic. It is a materials problem, a computational problem and a representation problem at once, which is why, perhaps, it has resisted an easy solution for so long.
If you work in games and see any of this differently, I would genuinely like to hear from you.
Contact me:
Beth Thompson, Founder, Lux Symbolica SASU, Paris. IATSE Local 706, Makeup Artists and Hair Stylists Guild. © 2026 Lux Symbolica SASU
References
Research
Kim, T., Schor, A., Posada, J., & Menon, A. V. (2026). The Racial Character of Computer Graphics Research. Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT '26), 1946–1978. doi:10.1145/3805689.3812275
Kim, T., Rushmeier, H., Dorsey, J., Nowrouzezahrai, D., Syed, R., Jarosz, W., & Darke, A. M. (2022). Countering Racial Bias in Computer Graphics Research. SIGGRAPH '22 Talks. arXiv:2103.15163
Shi, A., Bertails-Descoubes, F., Darke, A. M., & Kim, T. (2026). Curvature Space Editing of Highly-Coiled Hair. SIGGRAPH Conference Papers '26. doi:10.1145/3799902.3811115
Wu, H., Shi, A., Darke, A. M., & Kim, T. (2024). Curly-Cue: Geometric Methods for Highly Coiled Hair. SIGGRAPH Asia 2024 Conference Papers. doi:10.1145/3680528.3687641
Shi, A., Wu, H., Parr, J., Darke, A. M., & Kim, T. (2023). Lifted Curls: A Model for Tightly Coiled Hair Simulation. ACM SIGGRAPH/Eurographics Symposium on Computer Animation (SCA '23). doi:10.1145/3606920
Pernin, G., Davison, M., & Andrus, C. (2026). Practical Methods for Kinky-Coily Hair in Animation Production. SIGGRAPH '26 Talks. Netflix Animation Studios. doi:10.1145/3799818.3812100
Bertails, F., Audoly, B., Cani, M.-P., Querleux, B., Leroy, F., & Lévêque, J.-L. (2006). Super-Helices for Predicting the Dynamics of Natural Hair. ACM Transactions on Graphics, 25(3), 1180–1187. doi:10.1145/1141911.1142012
Wang, Z., Nam, G., Stuyck, T., Lombardi, S., Cao, C., Saragih, J., Zollhöfer, M., Hodgins, J., & Lassner, C. (2023). NeuWigs: A Neural Dynamic Model for Volumetric Hair Capture and Animation. CVPR 2023, 8641–8651. arXiv:2212.00613
Lin, G., Larionov, E., Chen, H., Roble, D., & Stuyck, T. (2026). Neuralocks: Real-Time Dynamic Neural Hair Simulation. Computer Graphics Forum. doi:10.1111/cgf.70407
Quaffure: Real-Time Quasi-Static Neural Hair Simulation (2024). Preprint. arXiv:2412.10061
Strands2Cards: Automatic Generation of Hair Cards from Strands (2025). SIGGRAPH Asia 2025 Conference Papers. doi:10.1145/3757377.3763864
Auto Hair Card Extraction for Smooth Hair with Differentiable Rendering (2025). Preprint. arXiv:2505.18805
Industry presentations, documentation and press
Strand-based Hair Rendering in Frostbite (2019). SIGGRAPH Course: Advances in Real-Time Rendering in Games.
Probe-based Lighting, Strand-based Hair System, and Physical Hair Shading in Unity's "Enemies". SIGGRAPH Course: Advances in Real-Time Rendering in Games.
Strand-based Hair and Fur in "Indiana Jones and the Great Circle". SIGGRAPH Course: Advances in Real-Time Rendering in Games.
AMD (5 March 2013). Announcement of TressFX Hair in Tomb Raider, with Crystal Dynamics. Reported by Overclockers.
NVIDIA (2015). The Witcher 3: Wild Hunt Graphics, Performance & Tweaking Guide. nvidia.com
Epic Games (2019). An Early Look at Next-Generation Real-Time Hair and Fur. unrealengine.com
CG Channel (December 2020). 10 Key Features for CG Artists from Unreal Engine 4.26. cgchannel.com
Epic Games. Groom Components and Assets in Unreal Engine (Unreal Engine 5.8 documentation). dev.epicgames.com
NVIDIA (1 September 2026). NVIDIA DLSS 5 Available Now: 3D-Guided Neural Rendering Debuts in NBA 2K27. nvidia.com
IEEE Spectrum (September 2026). Neural Rendering Redefines Photo-real Faces in Games. spectrum.ieee.org
Tech Xplore (20 July 2026). With Help from the Cosmetics Field, Animators Figure Out the Math of Highly Coiled Hair. techxplore.com
StatPearls. Anatomy, Hair. NCBI Bookshelf NBK513312. ncbi.nlm.nih.gov
Comments