ACM SIGGRAPH 2026

LightOpt: Lights Optimization for Real-time Rendering

A differentiable lighting optimization pipeline that reduces active lights in real-time game scenes while preserving the rendered appearance.

Tuo Chen¹, Luyan Cao², Kui Wu², Shimin Hu¹

¹Tsinghua University ²LIGHTSPEED

Watch video Paper PDF BibTeX

Input: 121 lights

Ours: 57 lights

5× difference

LightOpt reduces the Valley scene from 121 to 57 lights while preserving appearance, reducing average light overdraw from 2.40 to 1.67 in this view.

Abstract

Modern game scenes often require hundreds of lights to achieve rich visual detail, yet evaluating many lights per pixel remains a major performance bottleneck in real-time rendering, especially on low-end platforms. We present a differentiable lighting optimization pipeline that automatically reduces the number of active lights while preserving the final rendered appearance. Our method jointly optimizes light parameters and dynamically performs light removal, merging, and insertion during optimization. We introduce two loss functions to minimize redundant light overlap and prevent under-illumination, and extend our approach to support Time-of-Day lighting. Experiments on production game scenes show that our method reduces lighting cost by 18–52% while consistently outperforming artist-authored light reductions in visual fidelity.

18–52%lighting cost reduction

20–53%active light reduction

2.07×maximum shading speedup

10 minoptimization time per scene

Overview Video

The project video summarizes the motivation, pipeline, and rendering results of LightOpt.

Method Overview

LightOpt optimizes lights as a preprocessing step and can be integrated into existing tiled, clustered, or tiled-Z rendering pipelines.

Irradiance-field target

Surface samples are drawn from the scene, and the original irradiance field is used as a view-independent optimization target.

Overlap and attraction losses

The overlap loss discourages redundant light coverage, while the attraction loss pulls nearby lights toward under-lit regions.

Dynamic light controller

LightOpt alternates differentiable parameter optimization with light removal, merging, and insertion to adapt the light set.

Surface samples for irradiance optimization — Scene surfaces are discretized with Poisson disk samples to evaluate the irradiance field.

Extended light range for differentiable optimization — Extending the light range enables differentiable behavior as lights move across influence boundaries.

Results

LightOpt is evaluated on production game scenes with indoor, outdoor, and Time-of-Day lighting setups.

Tavern input lighting — Tavern: 210 input lights are reduced to 126 optimized lights with lower ALO and RMSE 0.041.

Tavern optimized lighting — Tavern: 210 input lights are reduced to 126 optimized lights with lower ALO and RMSE 0.041.

Cave input lighting — Cave: the optimized result reduces deferred shading time from 0.71 ms to 0.49 ms on RTX 3090.

Cave optimized lighting — Cave: the optimized result reduces deferred shading time from 0.71 ms to 0.49 ms on RTX 3090.

Square Time-of-Day input lighting — Square demonstrates the Time-of-Day extension, optimizing dynamic light parameters across keyframes.

Square Time-of-Day optimized lighting — Square demonstrates the Time-of-Day extension, optimizing dynamic light parameters across keyframes.

Highlights

Across five real scenes, LightOpt consistently lowers average light overdraw and deferred shading time while preserving illumination. The method also outperforms an industrial view-dependent light-pruning baseline in both speed and visual fidelity.

Tiled-Z rendering Deferred shading Time-of-Day

Scene	Input lights	Ours lights	Mean ALO ↓	Mean shading ↓	Mean RMSE ↓	Mean speedup
Valley	121	57	5.78 → 2.54	0.95 → 0.57 ms	0.046	1.67×
Tavern	210	126	3.35 → 1.72	0.71 → 0.52 ms	0.041	1.37×
Square ToD	314	249	3.28 → 1.57	0.83 → 0.56 ms	0.043	1.48×
Cave	80	64	3.43 → 2.05	0.71 → 0.49 ms	0.037	1.45×
Street	135	95	3.83 → 1.68	0.67 → 0.45 ms	0.033	1.49×

Citation

If you find this work useful, please cite the paper.

@inproceedings{chen2026lightopt,
  title     = {LightOpt: Lights Optimization for Real-time Rendering},
  author    = {Chen, Tuo and Cao, Luyan and Wu, Kui and Hu, Shimin},
  booktitle = {SIGGRAPH Conference Papers '26},
  year      = {2026}
}