01RESEARCH

Generative Lighting

Training AI on analog photography

ComfyUILoRAFilm Reference

The challenge with AI-generated imagery isn't capability—it's control. Current diffusion models can produce stunning results, but "stunning" means nothing in production when you need consistent, directable lighting that matches a specific creative vision.

This research focuses on training custom LoRA models using medium-format film references as ground truth. The hypothesis: if we can encode the characteristics of analog lighting—the way Kodak Portra renders skin, how Fuji Velvia handles saturation in shadows—we can create AI tools that respond to lighting direction the way a skilled cinematographer would interpret notes on set.

The goal isn't to replicate film aesthetics superficially. It's to create a controllable system where "warm, soft key with cool fill" produces predictable, reproducible results across batches.

Assistant

Ask about Generative Lighting...

Key Points

  • Custom LoRA training on curated film reference library
  • Deterministic lighting response to directional prompts
  • Batch-invariant inference for reproducible lighting generation
  • Integration with existing ComfyUI pipelines
USD Workflows