ControlHair: Controllable Hair Rendering

• ControlHair is a physics-informed framework that uses simulation and control-signal extraction to enable dynamic hair rendering and animation.
• It employs a cascaded pipeline that separates 3D hair simulation from video diffusion to achieve photorealistic and controllable outputs.
• The method enhances applications in avatar synthesis, interactive grooming, and robotics by using explicit control variables for hair behavior.

ControlHair denotes both a specific physics-informed video diffusion framework for controllable dynamic hair rendering and a broader class of methods that promote hair from a passive texture to an explicitly steerable geometric, dynamical, or compositional object in generative imaging, 3D avatar construction, interactive grooming, and robotic manipulation. In its canonical physics-informed form, the task is posed as conditional video generation,

$$\mathbf{V}=\mathcal{G}\big(\mathbf{I}_{\text{ref}},\mathcal{P}_{\text{hair}},\mathcal{F},\mathcal{T}_{\text{cam}}\big),$$

where a reference image, intrinsic hair parameters, external forces, and a camera trajectory determine a photorealistic video whose hair dynamics obey specified conditions (Lin et al., 25 Sep 2025). Related systems use simulator-derived strand maps, UVW strand descriptors, attention-derived hair masks, multi-view pose embeddings, or explicit strand/card parameterizations to make hair controllable across image editing, video animation, avatar synthesis, and physical manipulation (Chang et al., 11 Feb 2026, Jeong et al., 24 Jun 2026, Sun et al., 10 Jul 2025).

1. Conceptual scope

ControlHair addresses a persistent limitation of pose-driven image-to-video and portrait editing systems: hair is often treated as part of the head texture, so it either sticks to the skull, behaves like blurry cloth, or changes in ways that are not explicitly controllable. The central claim of the ControlHair line of work is that hair requires its own control variables, distinct from body pose, text prompts, or generic image conditioning (Lin et al., 25 Sep 2025).

In the specific framework titled "ControlHair: Physically-based Video Diffusion for Controllable Dynamic Hair Rendering" (Lin et al., 25 Sep 2025), the decisive design choice is a cascaded separation between physics reasoning and photorealistic rendering. A simulator produces per-frame 3D hair geometry under user-specified parameters such as mass, stiffness, damping, gravity, wind, human motion, and camera trajectory; those geometric states are then converted into image-shaped control signals consumed by a video diffusion model. This design rejects a direct parameter-to-video mapping. The paper states that real videos cannot easily be labeled with ground-truth physics parameters, and therefore trains the diffusion model on geometry-derived control signals rather than on raw physical inputs.

The broader literature uses the term in a more general sense. HairWeaver explicitly states that it is, in effect, a “ControlHair” system because it turns precise hair motion descriptors into realistic, controllable hair animation from a single image (Chang et al., 11 Feb 2026). H-Adapter describes ControlHair as controlling where and how reference hair affects the source image inside a diffusion pipeline (Jeong et al., 24 Jun 2026). Digital Salon uses the notion for a multimodal authoring environment in which text, GUI operations, and physics parameters jointly control 3D hair generation, simulation, grooming, and image synthesis (He et al., 10 Jul 2025). Across these variants, the unifying theme is the externalization of hair control into a representation that is spatially localized, semantically meaningful, or physically interpretable.

A common misconception is that text prompting or body pose conditioning alone suffices for hair control. The ControlHair evaluation argues the opposite: text is too coarse to capture per-frame dynamics and direction, while pose-only systems control head motion without encoding hair-specific motion, leaving hair behavior to the model prior (Lin et al., 25 Sep 2025).

2. Physics-informed video diffusion

The canonical ControlHair pipeline has three stages: simulator-based physics encoding, per-frame control-signal extraction, and conditioned video diffusion (Lin et al., 25 Sep 2025). In Stage 1, a 3D hair strand model is reconstructed from the reference image using DiffLocks, then animated with Blender’s particle hair system under intrinsic parameters $\mathcal{P}_{\text{hair}}$ and external forces $\mathcal{F}$. The resulting geometry sequence is

$$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$

The same stage also instantiates a rigged human model sequence $\mathbf{M}_{1:T}$, so head and body motion directly drive the simulated hair.

Stage 2 converts geometry into 2D controls. For each frame $t$, 3D hair and body are projected with the camera pose $\mathbf{T}_{\text{cam},t}$: $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$ From the projected frame, ControlHair extracts strand maps $\Phi_{\text{strand}}$, 2D human pose $\Phi_{\text{pose}}$, and a binary hair mask $\mathcal{P}_{\text{hair}}$0. The control image is then

$\mathcal{P}_{\text{hair}}$1

This construction assigns strand maps to hair regions and pose features to non-hair regions. The resulting sequence $\mathcal{P}_{\text{hair}}$2 has the same spatiotemporal shape as the target video and is deliberately independent of raw physics parameters.

Stage 3 uses a conditioned video diffusion model. Appearance is encoded from the reference image by a Wan 2.1 image encoder $\mathcal{P}_{\text{hair}}$3; the dynamic control sequence is encoded by a 3D convolutional encoder $\mathcal{P}_{\text{hair}}$4; and a 2D control encoder $\mathcal{P}_{\text{hair}}$5 processes the reference-image control map $\mathcal{P}_{\text{hair}}$6. The rendered video is

$\mathcal{P}_{\text{hair}}$7

The backbone is Wan 2.1’s DiT, trained with the standard diffusion objective

$\mathcal{P}_{\text{hair}}$8

with conditioning vector $\mathcal{P}_{\text{hair}}$9 aggregating appearance and control encodings (Lin et al., 25 Sep 2025).

This modularization is the defining architectural principle of ControlHair. Physics determines geometry; geometry determines control signals; diffusion determines photorealistic rendering. A plausible implication is that upgrades in simulation or control-signal extraction can be incorporated without redesigning the renderer, provided the interface remains a geometry-derived control map.

3. Conditioning schemes and control signals

Different ControlHair systems differ primarily in the representation used to expose controllable hair state to a generative backbone.

System	Control representation	Conditioning mechanism
ControlHair	Strand maps, 2D pose, camera-projected geometry	3D/2D control encoders into Wan 2.1 DiT
HairWeaver	Body motion maps and hair UVW maps	Motion-Context-LoRA and Pose Encoder
Stable-Hair v2	Polar-azimuth pose embeddings and temporal view sequence	Latent IdentityNet, Hair Extractor, temporal attention
H-Adapter	Attention-derived source-aligned hair edit mask	Mask-gated IP-Adapter inpainting

HairWeaver uses a training quadruplet

$\mathcal{F}$0

where $\mathcal{F}$1 is a UVW rasterization of hair strands whose channels encode root texture coordinates $\mathcal{F}$2 and normalized arc-length parameter $\mathcal{F}$3. Hair tokens are encoded by the same frozen VAE used for video latents and concatenated as context tokens; pose tokens are added as a residual modulation. Motion control is then learned by a Motion-Context-LoRA attached to a frozen LTX-Video DiT, while a temporary Sim2Real-Domain-LoRA is trained on CG videos and discarded at inference to preserve photoreal appearance (Chang et al., 11 Feb 2026).

Stable-Hair v2 addresses multi-view controllability rather than dynamics. It augments diffusion time embeddings with pose embeddings,

$\mathcal{F}$4

where $\mathcal{F}$5 and $\mathcal{F}$6 are polar and azimuth angles. The model combines a pose-controllable Latent IdentityNet, a Hair Extractor that injects reference-hair features through hair cross-attention, and temporal attention blocks that operate across the view sequence, enabling viewpoint-controlled and view-consistent hair transfer (Sun et al., 10 Jul 2025).

H-Adapter uses no explicit strand map. Instead, it induces spatially disentangled cross-attention by training with separate hair and non-hair objectives. Its masked reference-conditioned attention is

$\mathcal{F}$7

where $\mathcal{F}$8 is a source-aligned hair mask derived from attention itself. The hair-region loss and non-hair consistency loss explicitly force the reference branch to act only within hair regions, and the resulting attention maps are converted into masks for coarse-to-fine inpainting (Jeong et al., 24 Jun 2026).

HairDiffusion separates hairstyle and color conditioning. It introduces Multi-stage Hairstyle Blend,

$\mathcal{F}$9

so color information from a warped color proxy is injected once into the hair region, while hairstyle structure is maintained through ControlNet and a style proxy (Zeng et al., 2024). This makes multi-colored hair control a first-class operation rather than a side effect of prompt engineering.

4. Explicit 3D and compositional control spaces

A second major branch of ControlHair research encodes control directly in explicit 3D parameterizations. Perm formulates a parametric hairstyle model

$$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$0

where $$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$1 controls global hair shape through guide textures, $$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$2 controls local high-frequency residual textures, and $$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$3 specifies root positions on the scalp. Its central idea is a PCA-based strand representation in the frequency domain that disentangles low-frequency global structure from high-frequency curl and frizz, enabling separate manipulation of silhouette and local strand styling (He et al., 2024).

HairGPT turns hairstyle synthesis into a strand-token autoregressive problem. Each guide strand is parameterized as

$$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$4

with root UV coordinates $$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$5, coarse backbone $$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$6, and high-frequency style residual $$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$7. The model applies spatial decoupling across eight scalp regions and structural decoupling across layout, coarse, and style modes, using region-aware text embeddings and discrete geometric tokens to support compositional editing, synthesis of rare hairstyles, and adaptation to stylized domains (Luo et al., 9 May 2026).

CHARM performs an analogous decomposition for anime hair cards. Each control point stores five parameters—$$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$8, namely 3D position, width, and thickness—and hair cards are modeled as variable-length sequences of such control points. A transformer then autoregressively predicts card geometry and card termination tokens, enabling controllable generation from point clouds or images. Because mesh reconstruction from control points is deterministic, CHARM exposes direct card-level editing handles suitable for stylized hair authoring (He et al., 25 Sep 2025).

HairCUP extends compositional control to relightable 3D Gaussian avatars. It explicitly separates hairless face and hair into distinct latent spaces, hypernetworks, and Gaussian decoders, while anchoring hair Gaussian translations to a bald head mesh,

$$\mathbf{H}_{1:T}=\mathcal{S}\big(\mathbf{H},\mathcal{P}_{\text{hair}},\mathcal{F}\big).$$9

This enables hairstyle swapping across subjects while preserving identity, and few-shot fine-tuning for unseen subjects from monocular captures (Kim et al., 25 Jul 2025).

GroomCap, although framed as high-fidelity prior-free capture rather than editing, is control-ready in a different sense. It reconstructs explicit scalp-rooted strands from a neural implicit hair volume and then refines them with chained Gaussians under photometric and orientation-field supervision, producing geometry that can be directly used for cutting, rendering, or downstream simulation (Zhou et al., 2024). HVH similarly uses tracked guide strands as explicit control scaffolds for neural volumetric primitives, allowing new hair configurations to be rendered by editing guide trajectories (Wang et al., 2021).

5. Simulation, authoring, and robotic backends

ControlHair is not confined to diffusion models. Several systems treat controllable hair as a simulation or authoring problem first, and only then as a rendering problem.

Neuralocks is a strand-level neural dynamics engine intended for real-time control. It encodes rest-shape geometry into a 32-dimensional latent code, conditions on root UV, lock-level neighborhood code, body shape, and joint-velocity history for the previous $\mathbf{M}_{1:T}$0 frames with $\mathbf{M}_{1:T}$1, and predicts strand deformation in canonical scalp space. Training is fully self-supervised through a physically motivated objective,

$\mathbf{M}_{1:T}$2

with stretch, bend-twist, body collision, self-collision, style-preservation, adhesion, and inertia components. The dynamic network is about 1 MB, runs in $\mathbf{M}_{1:T}$3 ms for $\mathbf{M}_{1:T}$4 strands and $\mathbf{M}_{1:T}$5 ms for $\mathbf{M}_{1:T}$6 strands, and is presented as a practical control backend for real-time avatars (Lin et al., 7 Jul 2025).

Digital Salon makes this control stack interactive. It organizes hair authoring into four stages: text-guided retrieval, real-time physics-based hair simulation, interactive grooming/refinement, and hair-conditioned image generation. Its simulation stage uses an Augmented Mass-Spring model running at $\mathbf{M}_{1:T}$7–$\mathbf{M}_{1:T}$8 ms per frame for $\mathbf{M}_{1:T}$9–$t$0k strands, while the final image synthesis stage uses ControlNet with Stable Diffusion 1.5 conditioned on Canny edges from the current 3D viewport (He et al., 10 Jul 2025). In this formulation, ControlHair becomes a multimodal interface in which natural language, brush operations, trimming, wind, gravity, and strand-growth parameters all act as control channels over the same hair asset.

DYMO-Hair extends the concept into robotics. It represents hair as a voxelized volumetric state

$t$1

with occupancy and local orientation on a $t$2 grid, learns an action-conditioned latent dynamics model over this space, and uses Model Predictive Path Integral control to plan combing motions toward a visual goal hairstyle (Zhao et al., 7 Oct 2025). Simulation experiments report an average of $t$3 lower final geometric error and $t$4 higher success rate than the state-of-the-art system, and real-world experiments show zero-shot transfer to wigs on unseen hairstyles. This suggests that ControlHair can also denote physical control of real hair through learned world models rather than only generative rendering.

6. Applications and empirical evidence

The physics-informed ControlHair framework is evaluated on reconstruction from control signals extracted from ten Google VEO 2 videos. Averaged across the set, Wan 2.1 attains PSNR $t$5, SSIM $t$6, and LPIPS $t$7; UniAnimate-DiT attains PSNR $t$8, SSIM $t$9, and LPIPS $\mathbf{T}_{\text{cam},t}$0; ControlHair attains PSNR $\mathbf{T}_{\text{cam},t}$1, SSIM $\mathbf{T}_{\text{cam},t}$2, and LPIPS $\mathbf{T}_{\text{cam},t}$3 (Lin et al., 25 Sep 2025). The paper also reports strong-wind and head-motion examples in which text-conditioned or pose-only baselines fail to reproduce correct hair dynamics, whereas simulator-derived strand-map control visibly follows wind direction and inertial response.

HairWeaver reports stronger single-image hair animation under explicit motion descriptors. On its CG hair test set, it achieves hair-region metrics of approximately SSIM $\mathbf{T}_{\text{cam},t}$4, PSNR $\mathbf{T}_{\text{cam},t}$5, LPIPS $\mathbf{T}_{\text{cam},t}$6, FID $\mathbf{T}_{\text{cam},t}$7, and cd-FVD $\mathbf{T}_{\text{cam},t}$8, and on NeRSemble reports hair SSIM $\mathbf{T}_{\text{cam},t}$9, PSNR $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$0, LPIPS $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$1, and cd-FVD $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$2 (Chang et al., 11 Feb 2026). User studies in that work select HairWeaver roughly $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$3 of the time versus the best baseline at roughly $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$4 in one setting, and roughly $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$5 versus roughly $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$6 in a larger-scale setting.

For pose-robust hairstyle transfer, H-Adapter reports on a pose-different subset that includes yaw differences greater than $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$7. It achieves FID $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$8, $$\mathbf{I}_{\text{proj},t}=\Pi\big(\mathbf{H}_t,\mathbf{M}_t;\mathbf{T}_{\text{cam},t}\big).$$9 $\Phi_{\text{strand}}$0, and CLIP-I $\Phi_{\text{strand}}$1, with competitive SSIM $\Phi_{\text{strand}}$2 and PSNR $\Phi_{\text{strand}}$3, outperforming HairFastGAN, HairFusion, Stable-Hair, and HairCLIPv2 on the first three metrics under pose differences (Jeong et al., 24 Jun 2026). Stable-Hair v2 reports single-view CLIP-I $\Phi_{\text{strand}}$4, FID $\Phi_{\text{strand}}$5, PSNR $\Phi_{\text{strand}}$6, SSIM $\Phi_{\text{strand}}$7, and IDS $\Phi_{\text{strand}}$8, and in multi-view evaluation reports the best CLIP-I $\Phi_{\text{strand}}$9 and IDS $\Phi_{\text{pose}}$0 among the compared pipelines (Sun et al., 10 Jul 2025).

HairDiffusion reports text-driven editing metrics of IDS $\Phi_{\text{pose}}$1, PSNR $\Phi_{\text{pose}}$2, and SSIM $\Phi_{\text{pose}}$3, outperforming HairCLIPv2, HairCLIP, TediGAN, and DiffCLIP on that evaluation, while also showing improved FID, CLIP-FID, and SSIM in ablations for multi-color hair transfer when warping, PatchMatch, and bilateral filtering are retained (Zeng et al., 2024). HairCLIPv2, by contrast, demonstrates the strength of proxy feature blending for fine-grained local control: on CelebA-HQ it reports IDS $\Phi_{\text{pose}}$4, PSNR $\Phi_{\text{pose}}$5, and SSIM $\Phi_{\text{pose}}$6, and leads user studies for preservation in text-driven, sketch-based, and cross-modal editing (Wei et al., 2023).

In 3D avatar and geometry settings, HairCUP improves few-shot monocular avatar reconstruction over DELTA, reporting L1 $\Phi_{\text{pose}}$7, PSNR $\Phi_{\text{pose}}$8, SSIM $\Phi_{\text{pose}}$9, and LPIPS $\mathcal{P}_{\text{hair}}$00 on its eight-subject evaluation (Kim et al., 25 Jul 2025). CHARM reports the highest CLIP similarity, $\mathcal{P}_{\text{hair}}$01, among compared anime-hair generators and achieves the best or tied-best geometric metrics on AnimeHair (He et al., 25 Sep 2025). These results collectively indicate that ControlHair has become an empirical program rather than a purely conceptual one: the literature evaluates controllability through motion alignment, identity preservation, region-specific faithfulness, view consistency, and explicit geometry quality.

7. Limitations and future directions

Several limitations recur across the literature. First, accurate hair control remains representation-limited. ControlHair itself depends on 3D hair reconstruction from a single image and on Blender’s simulator, so reconstruction error or simulator inadequacy can propagate into control signals (Lin et al., 25 Sep 2025). HairWeaver depends on CG-derived UVW motion descriptors and a Sim2Real strategy that may still be sensitive to domain mismatch (Chang et al., 11 Feb 2026). H-Adapter notes difficulties with bangs placement, occlusions, and head coverings, especially when attention-derived masks no longer cleanly separate hair from non-hair (Jeong et al., 24 Jun 2026).

Second, many systems expose only partial disentanglement. Stable-Hair v2 offers pose/view control and reference-driven hairstyle transfer, but does not define explicit sliders for length, curliness, or color independent of the reference (Sun et al., 10 Jul 2025). HairDiffusion decouples hairstyle and color but still relies on accurate segmentation and a warping module that can fail under extreme pose or length mismatch (Zeng et al., 2024). Perm disentangles global and local geometry, but its learned prior remains bounded by the hairstyle coverage of USC-HairSalon and does not claim strong support for complex braids, tight Afro-textured hair, or wet hair (He et al., 2024).

Third, explicit 3D control spaces still face dataset and throughput bottlenecks. HairGPT notes that compact tokenization can smooth micro-features and that autoregressive synthesis takes roughly $\mathcal{P}_{\text{hair}}$02–$\mathcal{P}_{\text{hair}}$03 seconds per hairstyle (Luo et al., 9 May 2026). CHARM identifies repetitive cards, positional jumps, and coverage failures as typical autoregressive errors (He et al., 25 Sep 2025). GroomCap attains high fidelity but requires long per-subject optimization, and still struggles on very curly or extremely dark hair (Zhou et al., 2024).

The forward direction of the field is therefore dual. One direction strengthens the control substrate: better simulators, denser human-control signals, more accurate 3D hair reconstruction, richer semantic annotations, and improved segmentation. The other direction broadens the range of controllable outputs: HairWeaver explicitly suggests coupling with systems such as Neuralocks, Quaffure, and NeuWigs to close the loop between 3D simulation and 2D photorealistic rendering (Chang et al., 11 Feb 2026). HairCUP suggests a route toward compositional avatar priors with swappable hair and face components (Kim et al., 25 Jul 2025). DYMO-Hair shows that the same logic can extend beyond rendering to physical hairstyling by robots (Zhao et al., 7 Oct 2025). A plausible implication is that ControlHair will increasingly denote a family of interoperable modules—capture, simulation, control-signal extraction, generative rendering, and physical actuation—rather than a single algorithmic technique.