LightGen Explorer

An interactive web application exploring LightGen, the world's first all-optical generative AI chip, as described in the paper published in Science (December 2025).

🔗 Live Demo ← Update with your GitHub Pages URL

---

About the Paper

Citation

Chen, Y., Sun, X., Tan, L., Jiang, Y., Zhou, Y., Zhang, W., & Zhai, G. (2025). All-optical synthesis chip for large-scale intelligent semantic vision generation. Science, 390(6779), 1259-1265. DOI: 10.1126/science.adv7434

The Problem

Large-scale generative AI faces a severe computing power shortage:

• Stable Diffusion generates CO₂ emissions per 1000 inferences equivalent to driving 4.1 miles
• Llama-7B takes >3 seconds to infer 100 tokens on NVIDIA A10
• Energy and latency prohibit extensive edge deployment

The Solution: LightGen

LightGen is an all-optical photonic chip that performs generative AI computations using light instead of electrons.

Metric	LightGen	NVIDIA A100	Improvement
Computing Speed	3.57×10⁴ TOPS	~312 TOPS	>100×
Energy Efficiency	6.64×10² TOPS/W	~1.6 TOPS/W	>100×
Computing Density	2.62×10² TOPS/mm²	~0.4 TOPS/mm²	>100×

Key Innovations

1. 2.1+ Million Photonic Neurons - Integrated via 3D packaging in just 136.5 mm²
2. Optical Latent Space (OLS) - All-optical dimension conversion using single-mode fiber arrays
3. BOGT Training Algorithm - Bayes-based training independent of ground truth labels

Demonstrated Capabilities

• ✅ High-resolution (512×512) semantic image generation
• ✅ Image denoising with up to 20.4% noise reduction
• ✅ Style transfer (Van Gogh, Malevich, mosaic styles)
• ✅ 3D generation and semantic manipulation
• ✅ Video generation

---

About This Web Application

This interactive explorer helps you understand:

📖 Paper Overview

Summary of key findings, performance metrics, and comparison charts

🏗️ Architecture

3D visualization of the encoder → OLS → generator pipeline with animated light propagation

⚛️ Physics Simulation

Interactive photon simulation through diffractive metasurface layers with adjustable:

• Wavelength (400-700nm)
• Layer spacing
• Phase modulation

🔧 Manufacturing

Step-by-step fabrication process from digital training to 3D integration, with cost analysis

🌙 Night Vision Application

Explore how optical AI could enable ultra-low-latency, low-power night vision goggles

🏥 Medical Imaging Application

Interactive medical image enhancement with contrast, noise reduction, and edge enhancement controls

🚗 Autonomous Systems Application

Visualize the critical importance of processing latency for self-driving vehicles

📊 Parameter Explorer

Adjust system parameters and see computed performance metrics in real-time with 3D surface plots

---

Technologies Used

• Three.js - 3D graphics and physics simulations
• Chart.js - Interactive data visualizations
• MathJax - Mathematical equation rendering
• Vanilla JavaScript - No framework dependencies

---

Running Locally

# Clone the repository
git clone https://github.com/yourusername/repo-name.git
cd repo-name

# Start a local server (Python 3)
python3 -m http.server 8888

# Or with Node.js
npx serve .

Then open http://localhost:8888 in your browser.

---

Deploying to GitHub Pages

1. Push this repository to GitHub
2. Go to Settings → Pages
3. Under "Source", select Deploy from a branch
4. Choose main branch and / (root) folder
5. Click Save

Your site will be live at https://yourusername.github.io/repo-name/

---

File Structure

├── index.html      # Main HTML structure with 8 tabs
├── styles.css      # Dark theme styling
├── app.js          # Three.js visualizations & Chart.js graphs
├── README.md       # This file
└── science.adv7434.md   # Full paper content in markdown

---

Mathematical Formulations

The app includes key equations from optical computing:

Angular Spectrum Propagation: $$E(x,y,z) = \mathcal{F}^{-1}\left{ \mathcal{F}{E(x,y,0)} \cdot e^{i k_z z} \right}$$

Phase Modulation: $$E_{out}(x,y) = E_{in}(x,y) \cdot e^{i\phi(x,y)}$$

BOGT Training Loss: $$\mathcal{L} = D_{KL}(Q(Z|X) | P(Z))$$

Diffraction Limit: $$d_{min} = \frac{\lambda}{2 \cdot NA}$$

---

Future Possibilities

The LightGen architecture opens doors to:

• Fully analog optical systems - No digital conversion (e.g., optical night vision goggles)
• Edge AI - Ultra-low power inference on mobile devices
• Real-time medical imaging - Zero-latency image enhancement during surgery
• Autonomous vehicles - Sub-microsecond scene understanding

---

References

1. Chen, Y. et al. Science 390, 1259-1265 (2025)
2. Wetzstein, G. et al. Nature 588, 39-47 (2020) - Photonic computing review
3. Shen, Y. et al. Nat. Photonics 11, 441-446 (2017) - MZI computing
4. Mildenhall, B. et al. Commun. ACM 65, 99-106 (2021) - NeRF

---

License

This educational tool is provided for learning purposes. The original research is published in Science and is subject to their licensing terms.

---

Built to explore the future of AI computing at the speed of light ⚡

Preview