Microsoft Releases Lens: 3.8B-Parameter Text-to-Image Model Trained on 800M Image Dataset
Microsoft released Lens, a 3.8-parameter foundational text-to-image model trained on Lens-800M, an 800 million image-text corpus with GPT-4.1 captions. The model uses a 48-block MMDiT denoiser with FLUX.2 latents and supports generation up to 1440×1440 resolution across aspect ratios from 1:2 to 2:1.
Microsoft Releases Lens: 3.8B-Parameter Text-to-Image Model Trained on 800M Image Dataset
Microsoft released Lens, a 3.8-parameter foundational text-to-image model trained on Lens-800M, an 800 million image-text corpus with GPT-4.1 captions. The model uses a 48-block MMDiT denoiser architecture with FLUX.2 latents and supports generation up to 1440×1440 resolution.
Architecture and Training
Lens combines several technical approaches to achieve what Microsoft claims is competitive quality with "substantially less training compute than larger T2I models." The architecture uses:
- 48-block MMDiT (Multimodal Diffusion Transformer) denoiser
- FLUX.2 semantic VAE for latent representations
- Concatenated multi-layer GPT-OSS text features for prompt encoding
- Mixed-resolution training enabling flexible aspect ratios
The training dataset, Lens-800M, consists of 800 million image-text pairs with dense captions generated by GPT-4.1, which Microsoft describes as "maximizing information density per training batch."
Technical Capabilities
The model supports:
- Resolution range: up to 1440×1440 pixels
- Aspect ratios: 1:2 to 2:1
- Multilingual prompt following via GPT-OSS features
- Mixed-resolution inference across different aspect ratios
Microsoft also released post-trained variants including an RL-tuned version for improved visual quality and artifact suppression, and Lens-Turbo, a distilled variant supporting 4-step generation.
Release Details
The model is available on Hugging Face with minimal inference code for generating images from Lens DiT checkpoints. Training cutoff date was not disclosed. Pricing information, parameter count breakdown, and benchmark comparisons against established models like DALL-E 3, Midjourney, or Stable Diffusion were not provided in the release.
The sample gallery demonstrates capabilities across multiple styles including photorealistic scenes (landscapes, wildlife, architecture), artistic styles (oil painting, watercolor), text rendering (signs, typography), and multilingual prompts (French, Chinese references).
Project Team
The project is led by Dong Chen, Fangyun Wei, and Ziyu Wan, with core contributors including Jiawei Zhang, Jinjing Zhao, Sirui Zhang, Yang Yue, and Zhiyang Liang. The full team includes 24 researchers from Microsoft.
What This Means
Lens represents Microsoft's entry into the competitive foundational text-to-image model space, positioning efficiency as a key differentiator at 3.8B parameters versus competitors like Stability AI's SDXL (2.6B UNet parameters) or larger proprietary models. The use of GPT-4.1 for caption generation and FLUX.2's VAE suggests Microsoft is leveraging existing infrastructure for training data preparation. The lack of disclosed benchmarks, pricing, or training compute figures makes direct comparison difficult, though the 4-step Turbo variant suggests Microsoft is targeting real-time generation use cases. This release puts Microsoft in direct competition with Stability AI, Midjourney, and its Azure OpenAI partner in the text-to-image foundation model market.
Related Articles
OpenBMB's MiniCPM-V 4.7 35B-A3B MoE claims 13+ tokens/s on iPhone 17 Pro Max in 3 GB
OpenBMB has released MiniCPM-V 4.7 35B-A3B, the first mixture-of-experts model in the MiniCPM-V series, with open weights on Hugging Face. The company claims it runs at 13+ tokens/s using 3 GB of memory on an iPhone 17 Pro Max, and that it beats Qwen3.6 35B-A3B on overall score in less inference time in Instruct mode.
Microsoft releases Decision-1, a Qwen3.5-9B-based model for classification and routing, at $0.042 per 1M input tokens
Microsoft has released Decision-1, a decision model built on Qwen3.5-9B for classification, evaluation, and routing. Microsoft claims 83.5% accuracy across 36 benchmarks and 85 ms latency. Input tokens cost $0.042 per million, and output tokens are free.
Microsoft releases FrogNano-4B, an Apache 2.0 coding agent trained with RL on 1,500 synthetic tasks
Microsoft has released FrogNano-4B-2609, a repository-level coding agent derived from Qwen3.5-4B and published under Apache 2.0 with open weights. Microsoft says it was post-trained only with reinforcement learning on about 1,500 synthetic software-engineering tasks, with no stronger-model trajectories. It is evaluated at roughly 131K tokens of context.
Microsoft's Copilot gets access to local Windows files and OS-level actions under 'Hybrid Intelligence'
Microsoft announced an upgrade to Copilot at its Windows and Surface event that gives the assistant access to local files and the ability to take actions across Windows. The company calls the underlying approach "Hybrid Intelligence," which combines local and cloud AI models. Pricing, model details, and availability were not disclosed in the available reporting.
Comments
Loading...