Capturing materials: measure first, then generate

A material in a computer is a promise about light. It says that when light of a given colour arrives from a given direction, this much of it leaves in that direction. Get the promise right and a render, a camera simulation or a training image tells the truth about the surface. Get it wrong and everything downstream is confidently mistaken.

The last two years have been very good for the containers that hold that promise. Shared formats for materials and captured scenes have matured, neural methods have made complex materials cheap to render, and generative models can now paint a full physically based material onto a mesh in one pass. What none of this changes is where the numbers inside the container come from. For any use where the result has to match the world, they come from measurement.

The containers have converged

Physically based rendering, or PBR, describes a surface through a small set of parameters, base colour, roughness, metalness, normal, and more for coatings, sheen and transmission, fed into a model of how microscopic facets scatter light. For years every renderer had its own dialect. That is ending.

OpenPBR, a shading model hosted by the Academy Software Foundation as part of MaterialX, was designed to succeed Adobe's and Autodesk's standard surfaces. By early 2025 the MaterialX project listed OpenPBR 1.1 as the current release, available in MaterialX 1.39.2 and OpenUSD 25.02, which puts one surface description inside OpenUSD, the scene format used to exchange whole worlds.

Captured scenes have followed. 3D Gaussian splatting, which represents a scene as millions of small coloured, semi-transparent ellipsoids fitted to photographs, has spread from research into mapping, games and product capture since it was introduced in 2023. In 2025 Khronos, working with the Open Geospatial Consortium, Niantic Spatial, Cesium and Esri, added splats to glTF through experimental extensions, storing each splat as a point with position, rotation, scale, opacity and spherical harmonics. In February 2026 the baseline extension, KHR_gaussian_splatting, moved to release candidate, with CesiumJS, Esri ArcGIS, Scaniverse and XGRIDS among early adopters, and a viewer without splat support can fall back to drawing a point cloud.

Rendering got cheaper

The second shift is cost. At CES 2025 NVIDIA described RTX Neural Materials, which use a small network to compress the shader code of complex, layered materials normally kept for offline rendering, with material processing up to five times faster. The point is that a film-quality layered material, many lobes of reflection stacked on one another, can run at interactive rates.

Splatting has also grown out of its first limits. The original method assumed a perfect pinhole camera and handled only light that travels straight from the scene to the lens. 3DGUT, from NVIDIA and the University of Toronto at CVPR 2025, replaced the projection step so splats can be rendered exactly through distorted lenses and rolling shutters, and traced for secondary rays such as reflection and refraction, while keeping the speed of rasterisation. For anyone simulating a real camera, a fisheye on a ceiling or a wide lens on a vehicle, that matters.

Generation got good

The third shift is generation. Material Anything, at CVPR 2025, is a diffusion framework that produces full PBR materials for texture-less, albedo-only, generated or scanned meshes under varied lighting, trained on a dataset of over 80,000 objects with high-quality PBR materials. The results are clean and the maps are ready for a production pipeline.

It is worth being exact about what such a model learns. It learns what materials in its training set look like: their distributions of colour, roughness and normal detail, as artists and capture tools authored them. Its output is a plausible material for the object. Whether that material matches a particular real fabric, a particular painted steel or a particular batch of stone is a question the model has no access to, because nobody showed it that sample.

Where the light gets baked in

Capture has its own version of the problem. A splat scene fitted to photographs reproduces those photographs very well, and in doing so absorbs the light they were taken in. The shadows, the colour of the sky and the reflections become part of each splat's colour. A 2025 paper on inverse rendering with Gaussians names this directly: illumination baked into the representation makes editing and relighting a critical limitation. Several groups now separate material from light. RNG, at CVPR 2025, targets exactly the hard cases for fashion and products, objects with soft, ill-defined boundaries such as fur and fabric, and conditions each point's radiance on both the view and the light direction so they can be relit.

That work is moving fast, and it points the same way as everything above. A material separated from its lighting is a set of reflectance values, and those values are only as good as what they were fitted against.

Measure first

The order that keeps a material honest is therefore simple. Measure the real sample: how it reflects across angles and wavelengths, how its surface varies, and for cloth, how it bends, stretches and shears. Fit a representation to those measurements in an open format. Record the range the fit holds for, the angles, lights and scales at which it was checked. Then generate: variations, new colourways, the same weave in other yarns, new lighting. Generation is excellent at filling in around a measured anchor. Without the anchor it fills in around an average.

This matters most where rendered images become evidence: a camera plan judged on simulated views, a perception model trained on synthetic frames, a buyer deciding on a fabric from a screen. In each case a wrong roughness or a baked-in shadow is a wrong answer delivered at full confidence.

Where Livgens fits

Livgens represents the thing exactly, in sound, light, matter and motion, held as the thing itself with its measurements and the range they hold for, and Living Worlds runs it forward through its lifecycle. Livgens Render drives Blender, Unreal Engine, FreeCAD, KiCad and RealityScan, captures real places into models, and hands worlds over as OpenUSD and buildings as IFC. Its light and sight work shows what a camera sees from where it is mounted, with the truth behind every pixel.

In Render's events venue case study, each camera plan was tested at noon, golden hour and night, in rain and low sun and through the compressed recording, on each camera's simulated view, and every figure in the decision record is tagged by where it came from: the drawing, a photograph, a published source, a measurement on the twin or a simulation. What could not be measured was marked as assumed and pushed to both ends of its range. In the fashion house case study, prints and weaves are laid at their real repeat, and because no fabric has yet been measured on a bench, every drape is shown as unvalidated until it is scored against real cloth. Render is in early access.