PAJAMA / RELIGHT

ZERO TO HERO / NEXT CHAPTER

Move the light.
Understand what changes.

Your viewer can replay appearance. Now learn what it takes to change illumination: normals, materials, visibility and time. Eight short chapters connect the equations to a working Gaussian renderer.

Experiment with light →Start with the first principle
Captured RGB ≠ material→Known-material GPU lab→Inverse rendering + 4D

What runs today

1,536 synthetic Gaussian surfels · Lambert/GGX · five inspection modes · a moving light and deforming object · linear HDR composition. Uses the shared Rust/wgpu core. The current pretrained STG scene has no verified material fields; this lab uses a separate known-material asset.

A small idea at every step

01

A color is not a material

Separate camera motion, light motion and material changes.

02

Count the light, one direction at a time

Build the rendering equation from multiplication and addition.

03

Build a moving highlight

Read every factor in the actual GGX shader.

04

Send light through the splat pipeline

Follow two GPU passes and their color spaces.

05

Use SH for light, not just appearance

Understand lighting coefficients, transfer and TranSplat.

06

Learn materials without learning the wrong thing

Trace inverse rendering from observations to a relightable asset.

07

Add time without breaking the light

Derive inverse-transpose normals and dynamic visibility.

08

Bring relighting into the engine

Separate working capabilities from the next implementation milestones.

Where to study

Follow this sequence: foundations → environment lighting/BRDF → our GPU lab → one paper reproduction. These are public courses, books and implementation notes; none is presented as a dedicated Coursera-style 4DGS certification.

GAMES101 · foundations and radiometry

Chinese lectures, English slides. Review shading, ray tracing and radiometry; the official page links recordings.

GAMES202 · environment lighting, PRT, microfacets

Prioritize environment lighting 1–2 and high-quality shading 1, then shadows. A rendering course, not a GS certificate program.

PBRT 4 · light transport equation

Read after chapter 2; identify the physical quantity behind every symbol. A free online book, not a video course.

PBRT 4 · microfacet derivation

The complete distribution, masking and geometry behind the compact shader.

Filament · PBR engineering notes

Implementation reference for roughness, Fresnel, IBL and linear color; richer than this teaching subset.

RESEARCH SNAPSHOT / 2026-09-21

What has changed recently

A selective reading map, not an exhaustive ranking. Publication labels follow the official pages. “Code available” means a released implementation was observed, not that we reproduced its results. Availability can change.

Static inverse rendering

GS-IR · CVPR 2024 ↗

Geometry reconstruction, occlusion baking, then material/lighting estimation. A staged training baseline.

Official code available

Materials + visibility

Relightable 3D Gaussian · ECCV 2024 ↗

Material/lighting decomposition and ray-based visibility in Gaussian reconstruction.

Official code available

Illumination decomposition

GS-ID · ICCV 2025 ↗

Material priors plus environment/local illumination modeling: lighting need not be distant.

Code release pending per repository; README/assets observed

Volumetric video

BEAM · ACM MM 2025 ↗

Tracked Gaussians with PBR attributes, material estimation and visibility for dynamic relighting.

Paper verified; runnable official release not verified here

Head avatars

BecomingLit · NeurIPS 2025 ↗

Hybrid neural shading with FLAME and light-stage capture. Specialized data/model assumptions matter.

Official code; separate model/data access requirements

SH radiance transfer

TranSplat · ICCP 2026 ↗

Transfers trained 2D Gaussian surfels using source/target HDR lighting and normal-dependent irradiance ratios, with separate specular treatment.

Official code released July 2026

Deformable assets

DR-GS · ECCV 2026 ↗

Decoupled materials/light on 2D Gaussians, driven by simulation/animation deformation; distinct from learning arbitrary video motion.

Paper/project available; no code link observed on project page

Choose the next practical branch

For the engine: reproduce one static material asset with GS-IR or Relightable 3D Gaussian. For your SH studies: inspect TranSplat and its trained 2DGS + HDR input contract. For 4D: first prove that normals and visibility follow deformation, then study the different assumptions of BEAM, BecomingLit and DR-GS.

See the implementation milestones →