Microsoft Research's Lens achieves scale efficiency with detailed captions instead of scale
Microsoft Research's Lens proves detailed captions matter more than raw scale for training efficient image generators
The Decoder
Summary
Microsoft Research unveiled Lens, a 3.8-billion-parameter text-to-image model that matches larger competitors on benchmarks while using less training compute. The key was 800 million detailed captions generated by GPT-4.1, which proved more effective than vague web alt-text.