Summary:
Researchers in the UCLA Department of Electrical and Computer Engineering have developed an optics-based system to rapidly generate synthetic images. Background: Generative AI models that create synthetic images, human-like natural language processing capabilities, and even new protein designs are critical for a diverse array of applications. Well-known systems like ChatGPT, DALL-E3, and Sora have seen widespread adoption in many industries due to their versatility. However, as these models grow larger, their increasing power and memory consumption have become a concerning bottleneck. To overcome these emerging limitations and environmental concerns, there is an urgent need to develop new design methodologies to create scalable and power-efficient generative AI models.
Innovation:
Researchers led by Professor Aydogan Ozcan have developed optical generative models that can rapidly create synthetic images in an energy-efficient manner. By mapping random noise into phase patterns to seed image generation, an optical decoder can be used to create unique images. This method improves computational speed by approximately 3-fold when compared to conventional generative AI models. The team has demonstrated the ability to make both monochrome and multicolor images. They validated that the generated images were statistically comparable to pre-existing image datasets of handwritten digits, fashion products, human faces, and other image types. As optical generative models continue to rise in popularity, this technique presents an efficient and scalable solution for novel synthetic image generation.
Potential Applications:
- Synthetic image generation
- Augmented/virtual reality systems
- Animation
Advantages:
- Energy efficient
- Reduced computational load
- Ultra-fast processing
- Improved inference speed
- Scalable AI model
Development-To-Date:
Researchers have built the optical system and generated monochrome and multicolor synthetic images.
Related Papers:
Chen, S., Li, Y., Wang, Y. et al. Optical generative models. Nature 644, 903–911 (2025). https://doi.org/10.1038/s41586-025-09446-5
Reference:
UCLA Case No. 2025-067
Lead Inventor:
Professor Aydogan Ozcan, PhD, UCLA Department of Electrical & Computer Engineering