Can pre-integrated neural fields render complex emitters in real time?

NEF precomputes illumination, not just emitted radiance

The paper's central move is to change what gets precomputed. Earlier proxy representations store the radiance leaving an emitter as a 4D near-field proxy, which removes the cost of sampling paths inside the emitter assembly but still forces every receiver point to integrate over that proxy at runtime. NEF instead precomputes the pre-integrated outgoing direct illumination with respect to a receiver point, meaning any shaded surface point in the volume around the emitter.

That quantity is represented as a neural field parameterized by the receiver's position, normal, view direction, and material parameters, and learned by a network conditioned on a material model. Because the receiver's BRDF is folded into the learning objective, the integration is shifted into a one-time precomputation phase, and a single network evaluation replaces the entire integral over the proxy emitter.

A two-headed network returns noise-free unoccluded illumination

The architecture splits into diffuse and glossy heads, which lets one network serve both receiver types. At render time the pipeline first runs a G-buffer pass to obtain position, normal, outgoing direction, and roughness at each shading point; those channels are transformed into the emitter's local frame, encoded, and fed to the network to produce the outgoing radiance contribution pre-integrated with respect to the emitter.

The result is noise-free unoccluded direct illumination from a single network evaluation per shading point. This is a different guarantee from sampling-based estimators: Monte Carlo methods, including reservoir resampling, support arbitrary emitters but leave residual variance that typically requires a denoiser, which can introduce temporal and spatial artifacts. NEF avoids runtime sampling entirely and moves the computational burden to a one-time per-emitter offline training phase.

Training in the emitter's local frame makes the field portable

Training is performed in the emitter's local coordinate frame, and world-space queries are mapped into that frame at shading time rather than baked into the network weights. As a result, a trained NEF can be translated, rotated, and uniformly scaled at runtime without retraining, functioning as a portable lighting asset reusable across scenes under rigid transforms.

The paper frames the economics of this directly: a single training run amortizes across every scene and every instance in which the emitter is placed, much like a baked texture or precomputed BRDF. Internal interreflections, self-occlusion, spatially varying emission, and deformations are absorbed into the learned representation at zero additional runtime cost, because the training phase relies solely on path-traced radiance estimates for training data and the same procedure covers simple planar polygons, complex emissive meshes, and deformable emitter assemblies.

Accuracy matches analytical references on planar emitters

For diffuse receivers under uniform emission, the paper reports that NEF matches the analytical reference, confirming accuracy on the case analytical integration handles. For glossy receivers, the larger capacity of NEF allows it to achieve lower rendering error across various roughness values compared with Linearly Transformed Cosines, which the authors attribute to LTC's remaining approximation error.

One regime stays difficult: the near-mirror case where roughness approaches zero remains challenging for a finite neural representation, and the paper points to an example of increased error at lower roughness in the supplemental material. That boundary matters because it is exactly where a glossy analytical model is most nearly exact.

Inference cost tracks the emissive field, not polygon count

Analytical methods are extremely fast for low-polygon emitters, but their cost scales linearly with the number of emitter vertices, which becomes prohibitive for high-polygon emitters in real time. The paper reports a crossover point between the analytical methods and its glossy NEF at 267 emitter vertices, with the comparison shown in the supplemental material.

NEF's inference cost is instead determined by the network's size rather than the emitter's polygon count. The authors note that NEF inference does scale with emissive field complexity, since a highly spatially varying emitter benefits from a larger or more expressive network, but that this complexity is typically much lower than the geometry's: naively subdividing an emissive mesh raises its vertex count while leaving its emissive field largely unchanged, so the same NEF capacity suffices regardless of tessellation. All results in the paper share a single network configuration sized for the most demanding case, the Pierced Sphere, whose housing interreflections produce the hardest signal.

Storage, training time, and the visibility split

All results were generated with a single NEF architecture totaling 15 MB per emitter, or 30 MB for the glossy variant. For simpler emitters the authors describe this as often excessive and reducible. Training is a one-time offline cost ranging from approximately one minute for simple planar emitters to three hours for emitters with internal interreflections, dominated by generating path-traced training samples; the paper notes that reported training timings include the cost of path tracing because targets are generated on the fly rather than cached.

Visibility is handled separately from the learned field. Like existing analytical methods, NEF does not inherently capture occlusions of the emissive field caused by the surrounding scene. For simple emitters a single shadow ray toward the emitter center is often sufficient; for soft shadows, shadow mapping or a small ray-tracing budget combined with a denoiser applies directly. The paper reports that even 1 shadow ray per pixel yields perceptually accurate soft shadows when combined with the OptiX denoiser as a ratio estimator, with further samples producing diminishing returns. Crucially, shadow rays are only required against non-emitter scene geometry, because the NEF itself bakes in self-occlusion and internal interreflections within the emitter assembly.

Multiple emitters and sparse training signals limit the current scope

A scene with N emitters currently requires evaluating N independent NEFs and summing their contributions, unless the emitters are close enough to be represented jointly by a single NEF, so cost scales linearly with N. The authors see two mitigation avenues: a shared decoding MLP that learns to disentangle latents from multiple emitters, though a naive sum-then-decode scheme retains linear cost for latent fetching, and combining NEF with lighting hierarchies for sub-linear scaling.

Training data generation is the other pressure point. Targets are produced by path tracing the emitter assembly with next-event estimation. For emitters with housing geometry more intricate than those presented, heavily enclosing the emissive surface or admitting light only through narrow apertures, this strategy can produce a sparse training signal: most paths fail to reach the emitter and contribute zero, leaving the network to learn from a small fraction of informative samples. The paper suggests light tracing could reduce zero-radiance contributions, and that adaptively sampling the training loss could concentrate training data on high-frequency regions of the emissive field.

The 267-vertex crossover is the practical decision point

The most decision-relevant number in this paper is the crossover at 267 emitter vertices between analytical integration and the glossy NEF. Below it, analytical methods retain lower per-frame shading cost for simple polygons; above it, their linear scaling with vertex count makes NEF's geometry-independent cost the better trade. The paper's own framing supports reading the choice this way: generality and computational efficiency become especially relevant when extending beyond planar geometry, and the reported timings reflect the shading step only, measured at 1920×1080 on an RTX 4090 with a standardized architecture rather than per-emitter tuning.

Two conditions qualify that threshold. The near-mirror regime at low roughness remains a weak point for the finite neural representation, so a glossy analytical model may still be preferable there. And the crossover concerns unoccluded direct illumination: scene-level occlusion still needs shadow rays or a denoised visibility budget, and each additional emitter in a scene adds an independent NEF evaluation. Within those bounds, the paper's priority is clear: pay a one-time per-emitter training cost, from about one minute to three hours, to buy a portable 15 MB or 30 MB asset whose runtime cost stops tracking emitter complexity.

Sources used in this answer