Microsoft Lens Text-to-Image

Overview

Lens comes from a Microsoft research project on training foundational text-to-image models more efficiently. Microsoft says the 3.8B-parameter model reaches competitive quality with substantially less training compute than larger text-to-image models. Its denoiser is a 48-block MMDiT in the latent space of the FLUX.2 VAE. Text conditioning concatenates features from several layers of the GPT-OSS language model, and Microsoft attributes stronger prompt following and multilingual generalization to this design.

Pre-training used Lens-800M: 800 million image-text pairs with long captions written by GPT-4.1, so that each training batch carries more information. Mixed-resolution training covers aspect ratios from 1:2 to 2:1 at resolutions up to 1440x1440.

The four samples on this page are 1440x1440 squares: a Vermeer-style oil portrait, an explorer's desk with a world map and brass sextant, elephants below Kilimanjaro, and rice terraces at sunrise.

What you can build

• Research and prototype images from long scene descriptions, photographic or painted.
• Test renders in tall 9:16 and wide 2:1 frames, to see how the model composes outside the square.
• Concept art where the prompt pins down materials, light and camera angle.
• Label and signage mockups with short lettering. Microsoft's gallery shows quoted text on a honey jar label, a flower shop window and a roadside gas station sign.
• Images from prompts in other languages. One gallery prompt, the Eiffel Tower at dusk, is written entirely in French.

Inputs

• Prompt (required): a text description of the image. The prefilled example describes a Renaissance noblewoman in a blue velvet dress in one long sentence.
• Resolution: a base size of 1440 (default) or 1024. At 1440, a square is 1440x1440, 4:3 is 1664x1248 and 3:4 is 1248x1664. At 1024, the same ratios give smaller images.
• Aspect ratio: 1:1 (default), 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 2:1 or 1:2.
• Seed: a whole number from 0 to 9,999,999, default 0.

Outputs

One image per run, sized by the base resolution and aspect ratio you chose.

Recommended settings

• Write prompts the way Lens-800M captions read: long and specific about subject, setting, materials, lighting, color and medium or camera style.
• Use the 1440 base for final images and 1024 for drafts while you settle the wording.
• Keep the seed fixed while editing the prompt, so differences between outputs come from your wording.
• Put text you want rendered in quotation marks and keep it to a few words, as the gallery prompts do.

Limitations

• Microsoft warns that some prompts can produce inaccurate or inappropriate images, and that the largely web-sourced training data may carry biases and uneven representation.
• Generated images can be misleading or raise copyright and IP concerns.
• Multilingual support is described in general terms, with no list of tested languages, so check prompts in your language before relying on them.

Safety & compliance

Microsoft states that Lens is for research purposes only, is not intended for product or service deployment, and should be used in controlled settings with human oversight. Downstream users are responsible for safeguards such as content moderation, validation and compliance checks before broader use. Microsoft says the training data was processed to remove clearly identifiable personal information and reduce harmful content where possible. The project is released under the MIT license.