OpenHumanBRDF: Human Material Dataset

• OpenHumanBRDF is a full-body human dataset providing UV-aligned, high-quality material maps including displacement and subsurface scattering for enhanced realism.
• The dataset employs a controlled per-class material parameterization for four key dielectric materials, derived from 147 high-resolution 3D scans.
• It underpins the HumanMaterial framework by enabling progressive training with a fixed PBR model to alleviate ill-posedness in human material estimation.

Searching arXiv for the cited paper and topic to ground the article. arxiv_search({"query":"id:(Jiang et al., 24 Jul 2025) OR OpenHumanBRDF OR \"HumanMaterial: Human Material Estimation from a Single Image via Progressive Training\"", "max_results": 5, "sort_by": "submittedDate", "sort_order": "descending"}) Retrieving the arXiv record for the dataset-defining paper. arxiv_search({"query":"(Jiang et al., 24 Jul 2025)", "max_results": 3, "sort_by": "relevance", "sort_order": "descending"}) OpenHumanBRDF is a human-material dataset introduced in connection with full-body human inverse rendering based on physically-based rendering. It is described as a higher-quality dataset constructed from scanned real data and statistical material data, with the explicit goal of alleviating the ill-posedness that arises when material maps are weakly constrained by rendering supervision alone. In addition to normal, diffuse albedo, roughness, and specular albedo, it provides displacement and subsurface scattering, with the stated purpose of enhancing rendering realism, especially for skin. The dataset is presented together with the "HumanMaterial" framework, which uses it for single-image human material estimation (Jiang et al., 24 Jul 2025).

1. Dataset scope and subject composition

OpenHumanBRDF contains 147 full-body human models acquired from RenderPeople, and these are described as all originally 3D scans (Jiang et al., 24 Jul 2025). The subject pool is gender balanced (≈50 % male, 50 % female), spans six ethnicities evenly represented: Asian, Black, Indian, Middle Eastern, White, Hispanic/Latino (≈16.7 % each), and covers an age range: 19–50 (to maximize adult diversity).

The dataset assumes that the human body is composed of four principal dielectric materials, following Akenine-Möller et al. (RTR 2018):

• Hair
• Skin
• Fabric (clothing)
• Leather (accessories such as belts, shoes)

Diversity is described along several axes: body shape, clothing style and material (fabric vs leather), hair style, and skin tone. Each subject is uniformly rendered under multiple camera views and environment maps to cover a broad space of appearance (Jiang et al., 24 Jul 2025).

This construction places OpenHumanBRDF in a specific regime of human inverse rendering. It is not described as a generic material corpus spanning arbitrary object classes; rather, it is a controlled, full-body human dataset whose variability is organized around scanned geometry, UV-aligned maps, and four material classes.

2. Acquisition pipeline and material-map representation

The acquisition protocol begins with the high-resolution scanned meshes and UV layouts as provided by RenderPeople. In Blender (v2.8+), the pipeline assign[s] one of the four material classes to each mesh region, after which PBR parameters are set and six per-texel maps are baked via Blender’s Cycles/Eevee engines (Jiang et al., 24 Jul 2025).

The six baked maps are:

• normal (tangent-space)
• diffuse albedo
• roughness
• specular albedo
• subsurface scattering weight
• displacement

For each UV texel $(u,v)$, OpenHumanBRDF provides these maps in linear color space (32-bit float or 16-bit half float). Their semantics are specified as follows (Jiang et al., 24 Jul 2025):

• Normal (3-channel): world→tangent-space normal, mapped to $[-1,+1]$
• Diffuse albedo (3-channel): linear RGB reflectance
• Roughness (1-channel): scalar $\in[0,1]$
• Specular albedo (1-channel): scalar $F_0 \in[0,1]$
• Subsurface scattering weight (1-channel): scalar $\mathrm{sss} \in[0,1]$, used in Disney’s “BSSRDF” approximation
• Displacement (1-channel): signed distance from the base mesh surface along the normal direction, in meters

The derivation procedure is also explicit. In Blender’s Shader Editor, each principled BSDF is assigned the fixed $F_0$, roughness, and $\mathrm{sss}$ values, and the outputs of each parameter node are baked directly to image textures. Displacement is baked via Blender’s Displacement modifier into a height map (Jiang et al., 24 Jul 2025).

A central design feature is that all maps are baked to the subject’s UV space to ensure pixel-perfect alignment. This makes OpenHumanBRDF particularly suitable for supervised map regression, since appearance images, masks, and material parameters can be related through a stable UV parameterization rather than only through image-space correspondence.

3. Material parameterization and rendering model

The dataset uses a fixed per-class parameterization guided by RTR 2018 statistical ranges. Each of the four material classes is assigned scalar values for $F_0$, roughness, and SSS weight (Jiang et al., 24 Jul 2025).

Class	$F_0$	Roughness
Hair	0.239	0.50
Skin	0.184	0.40
Fabric	0.263	0.85
Leather	0.224	0.25

Class	SSS weight
Hair	0.00
Skin	0.08
Fabric	0.00
Leather	0.00

The parameter meanings are specified directly: $F_0$ is the Fresnel reflectance at normal incidence (i.e. specular albedo), roughness controls the microfacet distribution spread, and SSS weight blends a simple approximate subsurface scattering term in Disney BSDF (Jiang et al., 24 Jul 2025).

All appearance renders and render-loss supervision are based on a standard PBR shader with rendering equation

$[-1,+1]$0

where $[-1,+1]$1 is a surface point, $[-1,+1]$2 its normal, $[-1,+1]$3 the incident direction, $[-1,+1]$4 the exitant camera direction, and

$[-1,+1]$5

bundles the six material maps (Jiang et al., 24 Jul 2025).

The BSDF is specified as Disney’s BSDF (Burley 2012), described as a microfacet model with GGX normal distribution, Fresnel Schlick approximation, and an integrated subsurface term for skin. Its stated form is

$[-1,+1]$6

with $[-1,+1]$7 (Jiang et al., 24 Jul 2025).

This rendering model is important because OpenHumanBRDF is not only a collection of textures; it is tied to a concrete forward model for synthesis and supervision. The inclusion of displacement and subsurface scattering distinguishes it from simpler human-material datasets that only expose normal, albedo, roughness, and specular terms.

4. Lighting, viewpoints, calibration, and splits

OpenHumanBRDF includes 1 092 HDR environment maps (782 real from PolyHaven, 310 synthetic) (Jiang et al., 24 Jul 2025). For each subject, the camera is rotated on a fixed radius circle around the subject, and 100 azimuthal angles are sampled evenly to produce 100 views, corresponding to $[-1,+1]$8 increments.

For supervision generation, the pipeline renders the “appearance” (RGB) and all six material maps under two randomly selected environment maps. The split configuration is stated as follows (Jiang et al., 24 Jul 2025):

• Training split: 127 subjects → 127 × 100 = 12 700 data points
• Test split: 20 subjects → 20 × 10 (single initial view) = 200 data points

The statistical overview further specifies the sample contents:

• Train: 127 subjects, 12 700 multi-view samples (each sample = RGB appearance + mask + 6 material maps + 2 env-relit appearances)
• Test: 20 subjects, 200 single-view samples

Calibration and consistency are handled uniformly. The same camera intrinsics and world-to-camera transform procedure is used for all renders, and the UV-baking process ensures alignment across maps and rendered appearances (Jiang et al., 24 Jul 2025).

Preprocessing is also standardized. Each input appearance is accompanied by a foreground mask (from RemBG 2020 with optional alpha matting), and all maps are normalized to $[-1,+1]$9 or $\in[0,1]$0 where appropriate and saved in linear space (no gamma) (Jiang et al., 24 Jul 2025).

These design choices indicate that OpenHumanBRDF is intended for tightly controlled supervision. A plausible implication is that it supports both direct map estimation and render-based losses without requiring additional correspondence estimation between material maps and rendered observations.

5. Statistical structure and practical data organization

The dataset’s parameter distributions have a distinctive form. Because each material class has fixed roughness and specular values, the histograms of $\in[0,1]$1 and roughness exhibit four spikes at the class-specific settings. By contrast, diffuse albedos and normals are diverse due to per-subject color and geometry differences (Jiang et al., 24 Jul 2025).

This point matters for interpreting the dataset correctly. A common misunderstanding would be to treat OpenHumanBRDF as a source of continuously varying human BRDF measurements. The summary instead specifies a hybrid construction: geometric and appearance diversity come from scanned subjects and varied rendering conditions, while some BRDF parameters are discretized by material class. This suggests that the dataset emphasizes robust human material estimation under structured priors rather than unconstrained recovery of arbitrary continuous reflectance fields.

The storage format is designed for high-precision rendering workflows. The provided file formats are (Jiang et al., 24 Jul 2025):

• Appearance, mask, and baked maps: OpenEXR (.exr) or high-precision PNG (.png, 16-bit float) per channel
• Environment maps: HDR (.hdr) at 4 096×2 048 px
• Relit appearances: PNG/JPEG for quick preview, EXR for high precision

An example directory layout includes subjects.txt for split definitions, envs/ for the 1 092 HDR maps, per-subject folders with appearance/, materials/, masks/, and relit/, and within materials/ the files normal.exr, albedo.exr, roughness.exr, specular.exr, sss_weight.exr, and displacement.exr (Jiang et al., 24 Jul 2025). The example also lists relit/ as containing 5 novel-env relit images.

6. Role in HumanMaterial and methodological significance

OpenHumanBRDF is introduced in the paper together with HumanMaterial, a model for human material estimation from a single image via progressive training (Jiang et al., 24 Jul 2025). The paper motivates the dataset by noting that inverse rendering requires estimating multiple material maps and usually relies on rendering constraints, but the absence of constraints on the material maps makes inverse rendering an ill-posed task. Previous approaches are described as using material datasets with simplified material data and rendering equation, leading to rendering results with limited realism, especially that of skin (Jiang et al., 24 Jul 2025).

The stated contribution of OpenHumanBRDF within this framework is to provide stronger supervision through higher-quality material maps and a richer rendering model. Because the target space includes more maps, the paper argues that using an end-to-end model as in the previous work struggles to balance the importance among various material maps, and leads to model underfitting. HumanMaterial therefore first obtain[s] the initial material results via three prior models, and then refine[s] the results by a finetuning model. Since different maps have different significance for rendering results, the method introduces Controlled PBR Rendering (CPR) loss, which enhances the importance of the materials to be optimized during the training of prior models (Jiang et al., 24 Jul 2025).

In this context, OpenHumanBRDF functions as more than a static asset repository. It supplies aligned map supervision, relighting supervision, and a fixed PBR forward model that can support staged optimization strategies. This suggests a close coupling between dataset design and training methodology: the dataset’s structured map decomposition directly informs the progressive training regime.

7. Interpretation, strengths, and limitations

The strongest stated properties of OpenHumanBRDF are its use of scanned real data, statistical material data, pixel-perfect UV alignment, multiple camera views, and HDR environment-map illumination, together with explicit support for displacement and subsurface scattering (Jiang et al., 24 Jul 2025). These features make it well matched to full-body human inverse rendering pipelines that require both physically based supervision and map-level targets.

At the same time, the dataset summary makes several constraints explicit. The material taxonomy is restricted to four principal dielectric materials. The scalar parameters $\in[0,1]$2, roughness, and SSS weight are fixed per class rather than continuously sampled. The age range is 19–50, and subject identities are limited to the 147 full-body human models acquired from RenderPeople (Jiang et al., 24 Jul 2025). None of these are defects in themselves, but they delimit the dataset’s scope.

Access is described prospectively rather than as an already completed release. The authors state that they plan to open-source the dataset on GitHub (link to be released upon paper acceptance) and that it will be distributed under a Creative-Commons (CC BY-SA) license to allow academic and noncommercial use (Jiang et al., 24 Jul 2025).

Taken together, OpenHumanBRDF is best understood as a controlled, full-body human material dataset for inverse rendering under a Disney-style PBR model. Its main distinguishing property is not unconstrained reflectance diversity, but the combination of scanned human geometry, UV-aligned supervision, class-structured material parameterization, and rendering assets designed for relighting and material estimation.