Here is an observation that sounds obvious once said and has a real architectural consequence: subdividing a sphere does not change how it bounces light.
RADIANT — RADiance-based Inter-Actor Neural Transformer — was posted to arXiv on 23 October 2026 (arXiv:2610.32075) by Sarah Zhang, Zhiyang Dou, Minghao Guo, Benjamin Tod Jones and Wojciech Matusik. Its Comments field is empty and there is no journal reference, so this is an unreviewed preprint.
The setup
Feed-forward neural renderers are an interesting corner of rendering research: they “synthesize global illumination across novel 3D scenes without per-scene optimization or runtime ray tracing”. No training per scene, no path tracing at inference. You hand the network a scene it has not seen and it produces the lighting.
That is a different proposition from NeRF-family methods, which optimise per scene, and from real-time ray tracing, which does the physics at runtime on expensive hardware.
The existing transformer-based ones have a specific structural flaw, which the paper states cleanly:
Existing transformer-based renderers tokenize individual triangles, so self-attention scales quadratically with mesh resolution.
One token per triangle. Self-attention is O(N²) in tokens. So the cost of rendering a scene is quadratic in how finely its geometry happens to be subdivided.
The argument against triangles as tokens
And this is the part worth reading twice:
macroscopic radiative transport is governed by the underlying surface geometry and object arrangements rather than a particular mesh discretization; finer tessellation can thus inflate token counts without changing the underlying transport physics
Mesh density is an authoring artefact. It records how an asset was modelled, what the LOD pipeline did, and whether an artist hit subdivide twice — not anything about how light behaves. A triangle-tokenizing renderer therefore spends compute in proportion to a quantity that is irrelevant to its own answer. Two versions of the same scene at different tessellations should cost the same to light, and do not.
This matters practically because real asset libraries are wildly inconsistent in density. A scene assembled from several sources — exactly what OpenUSD libraries make easy — mixes a 200-triangle crate with a 400,000-triangle scanned engine, and the engine dominates the attention budget for no good reason.
The architecture
RADIANT is object-centric and works from point inputs. The central move:
By representing each object with a compact budget of 32 surface-patch tokens, RADIANT decouples attention sequence length from polygon counts.
Thirty-two tokens per object, regardless of geometry. Attention now scales with how many objects are in the scene — which is both a much smaller number and the number that actually governs inter-object light exchange.
A single network holds two parts:
- an inter-object scene transformer modelling 3D radiative exchange between objects
- a view decoder rendering images at target camera poses
That separation maps onto the physics: the lighting solution is view-independent, and a view is a query against it. Computing transport once and decoding multiple cameras from it is how offline renderers have always been structured, and it is pleasing to see it fall out of the architecture rather than be bolted on.
The abstract also begins describing how light transport is captured “without materializing O(N²) bias matrices” before the arXiv record truncates, so the mechanism there is in the paper and not something we will paraphrase from a fragment.
What is not here
No numbers. The abstract as published gives no render times, no quality metrics, no baseline comparisons. There is a clear architectural argument and no reported evidence that it wins, which with an empty Comments field means nothing has been reviewed either.
No code or project page appears on the arXiv record — only the abstract and the PDF. We checked; there is nothing to verify in either direction.
“32 tokens per object” is a budget, not a free lunch. A fixed representation per object necessarily discards detail, and the question any renderer like this has to answer is what it loses. Thirty-two surface patches can describe a crate well and a chandelier badly. High-frequency geometry that casts structured shadows — foliage, a lattice, a grille — is exactly where a coarse per-object summary should struggle, and the abstract does not say how it fares.
Object-centric means you need objects. A scene must be segmented into actors for this to apply. That is natural in USD or a game engine scene graph, and not natural for a single scanned mesh of an entire room.
Why it is worth watching anyway
The ideas most likely to survive in rendering are the ones that stop spending compute on quantities that do not affect the answer. Mip-mapping, level of detail, importance sampling and irradiance caching are all that same move.
“Attention should scale with scene composition, not with tessellation” is in that family, and the reason to flag an unreviewed preprint with no numbers is that the argument is checkable independently of this implementation. If it holds, every triangle-tokenizing neural renderer has a cost model tied to the wrong variable, and that is true whether or not RADIANT turns out to be the system that fixes it.
The author list includes Wojciech Matusik, whose MIT group has produced a long run of work in this area, which is a reason to expect the numbers to arrive rather than a substitute for them.
Sources
- RADIANT: Object-Centric Feed-Forward Global Illumination — arXiv:2610.32075
- arXiv PDF
- arXiv listing for cs.GR
- OpenUSD