Materials Are Becoming Learned Programs

For most 3D production, a material is built from familiar pieces: texture maps, parameters and shader code.

Base colour lives in one texture. Roughness and metallic values live in others. A shader graph defines how those values interact with light. More complex materials add layers, procedural masks, clearcoat, fibres or subsurface models until the renderer can reproduce the desired appearance.

SIGGRAPH 2026 research is pointing toward another representation.

VideoNeuMat extracts compact neural materials from video diffusion models. NVIDIA's richer-material work represents layered appearance with latent textures decoded by a shared MLP. 8DNA goes further by learning light transport through complex assets so effects such as subsurface scattering and glossy interreflection can be evaluated from a neural representation instead of recomputed from scratch.

I think materials are beginning to change from collections of authored parameters into small learned programs that answer a rendering question: given this point, view direction and lighting condition, what appearance should this surface produce?

A texture set stores values. A neural material stores behaviour.

Traditional PBR is powerful because the representation is explicit. Roughness, normals and metallic values have clear meanings, and the renderer combines them through a known model. That gives artists control, but it imposes a fixed vocabulary. Real materials can contain layered scattering, dust, coatings, fibres and other effects that are expensive to express with a small set of maps.

Neural appearance models take a different route. NVIDIA's 2024 Real-Time Neural Appearance Models work trained small networks to reproduce complex multilayered materials and compiled the learned representation into shader code for real-time rendering. The material could preserve effects that would otherwise require much heavier evaluation while still behaving like a shader inside the renderer.

The material is no longer only a bundle of maps. Part of its response to light is encoded in learned weights and latent data.

2026 makes the representation generative

The new work this year adds an important step: neural materials are becoming something generative systems can create directly.

VideoNeuMat starts from a video diffusion model that already contains broad visual knowledge learned from internet-scale data. The researchers finetune it to generate controlled material-observation videos under known camera and lighting trajectories. A second model then reconstructs a reusable neural material from 17 generated frames. That material can be rendered from new viewpoints and under new lighting.

This is different from generating a pretty image of marble. The output is an asset: it can be placed on geometry, relit and viewed from directions that were not part of the generated video. The appearance knowledge has been pulled out of the video model and converted into a representation a renderer can use.

The model becomes a material authoring source

Video generation models contain broad knowledge about how surfaces look across changing viewpoints and illumination, but that knowledge is entangled with geometry, camera motion and lighting.

VideoNeuMat treats the generative model almost like a virtual measurement device. By forcing it to produce controlled observations, the pipeline turns implicit visual knowledge into something closer to material measurements. The reconstruction model then converts those observations into a compact neural asset.

That suggests a new authoring pattern: instead of asking AI to generate the final render, use it to generate a reusable representation that remains inside the 3D pipeline.

I find that much more interesting.

PBR is not disappearing

PBR maps remain useful because they are interpretable and portable. A neural material is harder to inspect: if a highlight is wrong, there may not be one roughness slider responsible for it because behaviour can be distributed across latents and weights. I do not expect learned materials to replace standard PBR everywhere.

In fact, another line of research this year is trying to keep the two worlds connected. NVIDIA's procedural-data work begins with conventional PBR-style materials, expands them into richer layered appearance such as dust and clearcoat, then encodes the non-diffuse behaviour into a compact neural representation using latent textures and a shared decoder.

That hybrid direction makes sense. Keep the properties artists need to understand explicit. Learn the parts that are expensive or awkward to express directly.

A material can contain precomputed light transport

8DNA pushes the idea beyond local surface response.

Complex assets can contain appearance caused by light traveling through or around geometry: subsurface scattering, internal glossy reflections and fibre scattering. A conventional renderer has to simulate these paths, often with many samples.

8DNA learns an eight-dimensional light-transport representation from path-traced training samples. Once trained, the neural asset can reproduce expensive transport effects under changing near-field lighting with faster inference.

The boundary between "material" and "lighting solution" starts to blur. A learned asset can carry knowledge about how light moves through it, not only the static parameters used by a BRDF. For product visualisation, part of the response of translucent plastics, layered coatings or woven materials could live inside the asset itself.

This changes what optimisation means

With neural materials, optimisation can happen by distillation. Start with an expensive material, sample its behaviour, train a smaller network to reproduce the relevant response, then deploy that approximation instead of the original shader. This is already the logic behind neural appearance models and NVIDIA's neural-material direction.

That resembles software compilation more than texture baking: the source material describes high-fidelity intent, while training produces a target-specific runtime representation.

What this means for Blender workflows

For Blender, I would treat neural materials as another representation layer rather than a replacement for the shader editor.

During authoring, explicit nodes are still valuable because they are understandable. I can expose parameters to an artist, procedurally drive them and debug why the material behaves a certain way.

The learned representation becomes interesting at the boundaries: generation, deployment and capture. A material could be extracted from a prompt or generative model into a relightable asset, distilled from an expensive offline shader into a real-time neural shader, or reconstructed from observations without forcing every appearance effect into a fixed PBR parameter set.

The hard problem becomes editability

The production problem is control. Artists need local edits, versions and predictable outputs. A learned material that looks excellent but cannot answer "make only this coating slightly rougher" is difficult to use in production.

I expect future systems to separate interpretable controls from learned residual appearance. Some properties will stay as explicit parameters, while the network handles detail that would otherwise require large shader graphs or expensive scattering models.

We may end up authoring a material contract rather than every part of the material implementation.

The contract says what the artist controls. The learned representation solves more of how those controls become appearance.

My prediction

I think neural materials will become a normal asset type alongside texture-based and procedural materials.

Not every surface needs one. Simple painted metal is better served by simple parameters. Standard PBR remains useful because it is cheap, editable and portable.

The value appears when material behaviour is too rich for a small parameter set, too expensive for real-time evaluation, or difficult to recover manually from reference data.

What changed for me in 2026 is that the research is no longer only about compressing one expensive shader. VideoNeuMat shows that material knowledge can be extracted from generative models into reusable 3D assets. Richer-material work shows that generative systems can operate in neural material latent spaces. 8DNA shows that even complex light transport can become part of the learned asset representation.

The old material stack was maps plus code.

The next one may include weights, latents and learned light transport.

A material is starting to look less like a fixed set of textures and more like a small program trained to reproduce appearance.

Sources

1. NVIDIA Research, VideoNeuMat: Neural Material Extraction from Generative Video Models, SIGGRAPH 2026.
2. NVIDIA Research, Toward Richer Material Generation via Procedural Data Enhancement, SIGGRAPH 2026.
3. NVIDIA Research, 8DNA: 8D Neural Asset Light Transport by Distribution Learning, SIGGRAPH 2026.
4. NVIDIA Research, Real-Time Neural Appearance Models, SIGGRAPH 2024.
5. NVIDIA Research, An Introduction to Neural Shading, SIGGRAPH 2025.