Ternary Stable Diffusion

Apply ternary quantization to Stable Diffusion UNet for 15x compression

• Target: 4GB → 300MB model size
• Use case: Mobile and edge deployment
• Status: Research/Development

📊 Results Summary

See RESULTS.md for comprehensive proof of concept results and technical details.

🎉 Proof of Concept Results

First Ternary Image Generated (v0.1.0)

Compression Achieved:

• Original UNet (FP16): 327.8 MB
• Ternary UNet (2.5-bit): 21.9 MB
• Compression Ratio: 15.0x ✅

Technical Details:

• Model: Stable Diffusion v1.5
• Layers Converted: 282 (98 Conv2d, 184 Linear)
• Quantization: {-1, 0, +1} ternary weights
• Activation: FP16 (maintained)

Performance:

• Conversion Time: 4.4 seconds
• Generation Time: 33.6 seconds (10 steps, CPU)
• Device: CPU-compatible (no GPU required)

Deployment Potential:

• ✅ Mobile devices (22MB model size)
• ✅ Edge computing
• ✅ Privacy-first (on-device generation)
• ✅ Cost-effective (zero cloud costs)

Generated with ternary-quantized Stable Diffusion v1.5