Skip to main content
Process
research•January 10, 2025

Lighting-Aware Image Generation

Experiments in conditioning generative AI on explicit lighting direction and quality.

Most AI-generated images have flat, ambient lighting. This project explores whether we can teach models to respect lighting intent.

Approach

  • Key light direction (spherical coordinates)
  • Fill ratio
  • Color temperature
  • Specular intensity

**Current State** Promising results on portraits. Architectural scenes need more training data. The model definitely understands "light from upper left" now.

**Next Steps** Training on VFX-supervised data where ground truth lighting is known from the 3D setup.

AILightingControlNetResearch