Open 3.8B foundational text-to-image diffusion model from Microsoft. Reports competitive-with-or-better quality than FLUX and Stable Diffusion 3 at ~19.3% of Z-Image's training compute; the Lens-Turbo variant generates 1024² images in 0.84 s.

Three checkpoints released: base, Lens-Base, and Lens-Turbo. Microsoft's first open competitive frontier-class T2I model. Surfaced as a quiet HF org upload on May 22 alongside the arxiv submission — representative of exactly the release path the new HF-probe sweep step is designed to catch.

Model Details

Parameters 3.8B

Variants

Name Parameters Notes
Lens 3.8B
Lens-Base 3.8B
Lens-Turbo 3.8B Distilled fast variant; 1024² generation in 0.84 s

Paper

generationvisionopen-weightefficiency