MobileDiffusion: Rapid text-to-image generation on-device
Not a research paper. Google just shipped text-to-image generation that runs in half a second on your phone.
Why it matters
Google's MobileDiffusion solves a critical problem for on-device AI: running sophisticated generative models on consumer hardware. At 520M parameters and sub-second latency, this unlocks mobile deployment of AI features that previously required cloud infrastructure—shifting economics for companies building mobile-first products.
The key facts
8 to knowMobileDiffusion generates 512x512 images in 0.5 seconds on iOS/Android premium devices
Model size: 520M parameters (vs. billions for Stable Diffusion, DALL-E, Imagen)
One-step sampling via DiffusionGAN hybrid approach
Text encoder: CLIP-ViT/L14 (125M parameters)
VAE decoder achieves 50% latency improvement over Stable Diffusion baseline
Lightweight decoder: 9.8M parameters vs. SD's 49.5M, with better quality metrics (PSNR 30.2 vs. 26.7)
Published by Google Core ML team (Yang Zhao, Tingbo Hou)
Training via fine-tuning converges in <10K iterations using pre-trained diffusion weights
Go to the source
Google Research Blogblog.research.google
Publisher excerpt: Posted by Yang Zhao, Senior Software Engineer, and Tingbo Hou, Senior Staff Software Engineer, Core ML Text-to-image diffusion models have shown exceptional capabilities in generating high-quality images from text prompts. However, leading models feature billions of parameters and are consequently…