Microsoft Releases Lens-Turbo: 3.8B-Parameter Text-to-Image Model

Microsoft has released Lens-Turbo, a 3.8 billion-parameter text-to-image model trained on 800 million GPT-4.1-captioned images. The model uses a 48-block MMDiT (multi-modal diffusion transformer) architecture and supports generation at resolutions up to 1440×1440 pixels.

Technical Architecture

Lens combines several technical approaches:

• Training corpus: Lens-800M dataset containing 800 million image-text pairs with long-form GPT-4.1 captions
• Architecture: 48-block MMDiT denoiser with 3.8B parameters
• Latent encoding: Uses FLUX.2 semantic VAE for image encoding
• Text encoding: Concatenated multi-layer GPT-OSS features for prompt following and multilingual support
• Resolution handling: Mixed-resolution training enables aspect ratios from 1:2 to 2:1

Inference Speed

The distilled Lens-Turbo variant supports 4-step generation, according to Microsoft. The base model went through reinforcement learning post-training for improved visual quality and artifact suppression before distillation.

Resolution and Aspect Ratio Support

The model supports flexible output resolutions:

• Maximum resolution: 1440×1440 pixels
• Aspect ratio range: 1:2 to 2:1
• Multiple resolution presets: 1248×1664, 1664×1248, and square formats

Microsoft states the mixed-resolution training approach enables inference across different aspect ratios without quality degradation.

Training Efficiency Claims

Microsoft claims Lens reaches "competitive quality with substantially less training compute than larger T2I models" through dense-caption pre-training that maximizes information density per training batch. The company has not disclosed specific benchmark scores, training compute requirements, or comparisons to specific competing models.

Model Availability

The model is available on Hugging Face under the repository microsoft/Lens-Turbo. Microsoft has released minimal inference code for generating images from Lens DiT checkpoints. Pricing information for API access has not been disclosed.

What This Means

Lens-Turbo represents Microsoft's entry into the sub-4B parameter text-to-image model category, emphasizing training efficiency through high-quality captions rather than dataset scale. The 4-step distilled inference and flexible resolution support position it for applications requiring fast generation across varied aspect ratios. The reliance on GPT-4.1 for caption generation suggests Microsoft is leveraging its existing LLM infrastructure to improve training data quality, though the actual performance relative to models like Stable Diffusion 3 or FLUX.1 remains unverified without published benchmarks.