High-Efficiency Diffusion Models for On-Device Image Generation and Editing with Hung Bui - #753

What It Covers

Hung Bui explains how VinAI Research achieved efficient on-device AI by training smaller models that match larger model performance, developing one-step diffusion for real-time image generation, and building Vietnam's top AI research lab.

Key Questions Answered

• •Model Size Reduction: A sub-4-billion parameter Vietnamese language model outperformed the original 7-billion parameter version by iterating over the same dataset multiple times during training and applying minor optimization adjustments, proving smaller can be better with proper training techniques.
• •One-Step Diffusion: Swift Brush eliminates the typical 50-100 denoising steps in diffusion models by distilling multi-step knowledge into a single-step student network, achieving image generation in under 0.25 seconds while maintaining quality scores equal to or better than the original teacher model.
• •Image Editing Architecture: Swift Edit enables one-step image editing by training an inverted network that converts images to noise, then applies the one-step generation model. Training uses both real data and synthetic data generated by the efficient one-step model, creating highly intuitive loss functions.
• •Test-Time Scaling Advantage: Small models with inference-time scaling can outperform significantly larger models on specific tasks like math, making them viable for on-device deployment despite the increased compute requirements. This transforms the constraint of limited device resources into an opportunity for efficient specialized performance.