Microsoft has open-sourced Lens, a 3.8-billion-parameter text-to-image model family that rivals larger 6B+ models while being dramatically more efficient to train. In a peak BF16 TFLOPS normalized test (excluding caption regeneration), training Lens consumed only about 19.3% of the compute required by Alibaba's Tongyi Lab Z-Image model.
Data and architecture improvements are the secret sauce. The training set, Lens-800M, contains 800 million image-text pairs, but each caption is generated by GPT-4.1 and averages 109 words — much denser than typical short captions. The architecture uses 48 MMDiT blocks and a FLUX.2 semantic VAE. Text features come from GPT-OSS, with concatenated representations from layers 4, 12, 18, and 24 to boost prompt adherence and multilingual performance.
Microsoft offers three weight variants. The default Lens is fine-tuned with reinforcement learning (RL-tuned) and generates a 1024x1024 image in 3.15 seconds on a single NVIDIA H100 GPU over 20 steps. The distilled Lens-Turbo achieves the same resolution in just 0.84 seconds with only 4 steps. The base Lens-Base has no RL or distillation and runs 50 steps by default. The series natively supports any aspect ratio from 1:2 to 2:1 and mixed resolutions up to 1440x1440.
Weights are on Hugging Face in Safetensors and Diffusers formats under the MIT license, with inference code on GitHub. By combining high data density with extremely fast inference, Lens lowers the barrier for developers and academics to deploy large diffusion transformers.