Lens 3.8B model rivals 6B+ models with 80% less training compute

Microsoft researchers (Dong Chen, Fangyun Wei, and 19 others) present Lens, a 3.8B-parameter text-to-image model that matches or surpasses state-of-the-art 6B+ models while using only 19.3% of the training compute required by comparable models like Z-Image. The model's efficiency stems from two core strategies: maximizing data information density per batch and improving convergence speed through architectural choices.

First, training uses Lens-800M, a curated dataset of 800 million image-text pairs captioned by GPT-4.1 with an average of 109 words per caption—far richer than typical short captions. Each batch mixes multiple resolutions and aspect ratios to enlarge effective visual coverage per optimization step. Architecturally, Lens adopts a semantic VAE for better latent representations and a strong language encoder that accelerates optimization and enables multilingual generalization from English-only training data.

Post-training, the team applies reinforcement learning with taxonomy-driven prompts (Lens-RL-8K) and structured reward rubrics to suppress artifacts and improve visual quality. A reasoner module uses training-free system prompt search to better align user requests, and distillation accelerates generation to 4 steps. Thanks to its compact size, Lens generates a 1024² image in 3.15 seconds on a single NVIDIA H100 GPU, with the distilled turbo version achieving 0.84 seconds.

Lens generalizes to arbitrary aspect ratios from 1:2 to 2:1 and resolutions up to 1440², and supports prompts in multiple languages. This work shows that efficient training with high-quality data and smart architecture can outperform brute-force scaling, making high-quality image generation more accessible.