PHOTONIC TRANSPORT FIELD
A scene-compiled renderer whose nearest established relative is hierarchical radiosity, extended beyond diffuse patch exchange to carry directional reflection, refraction, and wavelength-dependent light.
compile the scene→retain transport→gather a view
Ray tracing is computationally expensive and fundamentally inefficient.
Light transport belongs to the scene, but a conventional probabilistic renderer repeatedly rediscovers it as sampled paths. Direct light, indirect light, reflection, and refraction are evaluated again across pixels, frames, and nearby camera views even when most of the illuminated scene has not changed. Acceleration structures make each geometric query cheaper; they do not make the repeated transport itself a retained part of the scene.
A retained scene-level light-transport representation can be discovered, solved, queried, and updated for less total work than tracing equivalent illumination path by path—even after charging the method for construction and maintenance.
Can sparse deterministic transport be performed more efficiently than nondeterministic probabilistic rendering by applying a transport-based radiance-field model?
Here, the radiance field is literal: light carried at a position, in a direction, and at a wavelength. The proposal is to compile the durable relationships of that field into the scene, solve illumination over those relationships, and let cameras query the result.
The implementation is hybrid. It does not abolish rays. It uses them to discover visibility, locate changing boundaries, and continue genuinely directional events; it avoids making a newly sampled complete light path the primary representation of every contribution.
The room, lights, materials, and solved diffuse field remain fixed while the camera completes this forty-second loop. View-dependent visibility, mirror returns, and glass transport are evaluated for each frame.
The same scene exposes colored diffuse transfer, indirect light in a recess, broad area-light shadows, mirror returns, and three-channel refraction through the prism.
These are demonstrations of the renderer, not separate methods. Each changes the geometry or lighting and contains 240 frames at 640 × 360, played for eight seconds at 30 fps.
Overlapping area lights · flat blockers
A violet side light and a white ceiling light shine through seven staggered fins. Partly visible emitters, overlapping shadows, and five small receivers stress exact subdivision of rectangular source visibility.
View-dependent reflection · repeated returns
Three differently oriented mirrors share a room with a cyan side light, two diffuse receiving panels, and a rough reflector. Reflected geometry changes with the camera while the scene's diffuse illumination remains retained.
Curved silhouettes · many partial occlusions
Twenty-eight suspended spheres and five colored flags create many curved silhouettes and partial occlusions. Their crossing contours stress camera-boundary detection and reconstruction.
Every result below was measured on a 2024 Mac mini, model identifier Mac16,10 and model number MU9D3LL/A, with an Apple M4 10-core CPU—four performance cores and six efficiency cores—and 16 GB of memory.
The figures separate what has been demonstrated from what remains a hypothesis. Playback rate is not render rate: the 30 fps videos are encoded presentations of frames that took much longer to calculate.
| Scene | Scene structure | Whole frame | Boundary work | Share |
|---|---|---|---|---|
| Aperture Canyon | 42 primitives · 230 graph links | 1.904 s | 0.999 s | 52% |
| Mirror Relay | 36 primitives · 154 graph links | 2.117 s | 1.146 s | 54% |
| Occlusion Garden | 63 primitives · 1,491 graph links | 1.263 s | 0.898 s | 71% |
In a separate fixed set of 4,096 spatial queries, bounding-volume traversal took 1.891 ms for 1,000 objects, 3.072 ms for 10,000, 3.398 ms for 100,000, and 5.253 ms for one million. Building the million-object index took 203.323 ms and its recorded primitive-plus-index storage was 380 MB. Those numbers measure geometric lookup, not a complete million-object illumination solve.
The retained field currently assumes fixed geometry, lights, and materials; changing them can invalidate transport, and an incremental update system has not been demonstrated. Geometry is limited to analytic rectangles, triangles, and spheres, so arbitrary skinned or deforming meshes, displacement, and fine texture-driven normal fields need another hierarchy. Diffuse transport is resolved at object or facet scale, which cannot preserve every fine indirect-light gradient.
The prism carries three wavelength channels rather than a continuous spectrum, rough directional transport is approximate, and participating media are absent. Fog, smoke, subsurface transport, continuous spectra, polarization, diffraction, and coherent wave interference would require transport state that this renderer does not contain.
Read the renderer, scene definitions, tests, and measurement receipts on GitHub