Agenda
Target-Driven Lighting Pipeline: Differentiable Light Optimization and AI-Assisted Authoring
Luyan Cao (Principle Graphics Engineer, LightSpeed Studios)
Kui Wu (Principal Scientist, LightSpeed Studios)
Pass Type:
Academic Pass, Festival Pass, Game Changer Pass, Indie Pass, Start-Up Pass -
Get your pass now!
Hot Topics:
AI
Track:
Programming, Visual Art
Audience Level:
Intermediate
Format:
Lecture
Vault Recording: TBD
Audience Level: Intermediate
This session presents a target-driven lighting pipeline that combines production light optimization with AI-assisted authoring. The first portion explains how the system analyzes light contribution, coverage, overlap, and local appearance; adjusts parameters; removes or consolidates inefficient lights; and creates replacements to preserve visual hierarchy within target-platform budgets. The second portion starts with AI-generated 2D lighting effect images. Artists review and approve the visual target before a renderer-in-the-loop fitter reconstructs an editable scene-lighting setup by optimizing light placement, color, intensity, range, and related parameters. The fitted results achieve a strong visual match and are already used by lighting artists in production. Through before-and-after visuals, measured data, and representative cases, the session shows how target definition, diagnostics, artist approval, and editable outputs make both optimization and AI-assisted creation reliable for daily use.
Takeaway
Attendees learn how to build a target-driven lighting pipeline with two integrated workflows: differentiably optimizing lights to meet platform budgets while preserving artist-approved appearance, and converting an approved, AI-generated 2D lighting image into an editable scene-lighting setup ready for daily production use.
Intended Audience
This session is for graphics programmers, rendering engineers, technical artists, lighting artists, and tools engineers working on real-time lighting pipelines. Familiarity with real-time lighting and basic GPU performance analysis is helpful; no prior machine-learning expertise is required.