Microsoft Lens Turbo · Text to Image

ABOUT THE WORKFLOW

Generate an Image from Text
Write a prompt describing the image you want. Lens Turbo generates it in 8 steps with strong prompt adherence, detailed scene composition, and multilingual text understanding. The output is a 1280x720 image. One prompt in, one image out.

Model

• Lens Turbo (3.8B) by Microsoft Research. A 48-block MMDiT denoiser paired with a GPT-OSS 20B text encoder and the FLUX.2 semantic VAE. Trained on 800 million densely captioned image-text pairs, it matches or outperforms 6B+ parameter models at 19.3% of their training compute. Supports aspect ratios from 1:2 to 2:1 and resolutions up to 1440x1440.

HOW IT WORKS

Step 1. Write your prompt
Describe the image: subject, setting, lighting, composition, and mood. Lens was trained on dense captions averaging 109 words, so it responds well to detailed, descriptive prompts. "An abandoned airport reclaimed by nature, ivy-covered airplanes, wildflowers across cracked runways, deer grazing nearby, soft pastel tones" gives the model clear direction.
Works great with: landscapes · architecture · nature · portraits · concept art · food photography

Step 2. Hit run and download
Lens Turbo generates the image in 8 steps and returns a 1280x720 result.
Ready for: Figma · Photoshop · Canva · social media · web · print

First time? Write a detailed prompt and hit run. Leave all settings as-is.

RECOMMENDED SETTINGS

Quick-start guide. Find the goal that matches yours and copy the settings.

• Standard image generation — 1280x720, 8 steps, CFG 1, euler sampler, seed fixed at 42. Write your prompt and run.

• Different aspect ratio — Change the latent width and height. 720x1280 for portrait. 1024x1024 for square. 1440x1440 for maximum resolution. Lens supports ratios from 1:2 to 2:1 natively.

• Faster generation — Drop steps from 8 to 4. Lens Turbo is distilled for 4-step generation and produces usable results at that count. Quality softens slightly.

• Explore different compositions — Change the seed from fixed to randomize. Each seed produces a different layout, angle, and interpretation of the same prompt.

• Multilingual prompts — Write prompts in languages other than English. The GPT-OSS text encoder generalizes across multiple languages despite English-only training data.

• Result looks generic or misses details — Write a longer, more detailed prompt. Lens was trained on captions averaging 109 words. It responds to specificity: lighting direction, material textures, spatial relationships, and color temperature all steer the output.

Prompt: Write like you are describing a scene to a photographer who cannot see it. "An abandoned airport reclaimed by nature, ivy-covered airplanes, wildflowers across cracked runways, deer grazing nearby, peaceful post-apocalyptic atmosphere, soft pastel tones" is specific. "Nice airport picture" is not.

LEARN

📹 Videos

• Intro to Floyo

• ComfyUI 101 Free Course ft. Sebastian Kamph

• Floyo 101 for Team Collaboration

✨ Quick links

• Help docs

• LoRA Training guide

• Discover more workflows

USE CASES

🎨 Concept Art and Scene Design
Generate detailed environment concepts, mood boards, and scene compositions from descriptive prompts with accurate spatial relationships and lighting.

📸 Editorial and Stock Photography
Produce photorealistic stills with natural lighting, material detail, and composition control for mood boards, pitch decks, and marketing assets.

📱 Social Media and Marketing Visuals
Generate on-brand images across aspect ratios for social posts, banners, and web content at sub-second speeds when using 4-step generation.

🌐 Multilingual Content Creation
Write prompts in multiple languages and generate images with consistent quality, making the model useful for international teams and non-English markets.

WHAT WORKS BEST / WHAT TO AVOID

✅ Works great

• Detailed, multi-sentence prompts with specific visual descriptions

• Nature, architecture, food, and environmental scenes

• Prompts that describe lighting, color, mood, and spatial layout

• Mixed aspect ratios from portrait to ultrawide

⚠️ May produce softer results

• Very short prompts with no visual detail (the model expects dense descriptions)

• Complex multi-character scenes with precise interaction requirements

• Exact text rendering inside images (not a primary strength)

• Prompts that rely on brand names or specific copyrighted references

FAQ

What is Microsoft Lens Turbo?
Lens Turbo is the distilled variant of Microsoft Research's Lens text-to-image model, released May 2026. At 3.8 billion parameters with a 48-block MMDiT denoiser, it matches or outperforms models with more than 6 billion parameters while requiring 19.3% of their training compute. The Turbo variant supports fast 4-to-8 step generation, producing a 1024x1024 image in under a second on an H100 GPU.

How is Lens different from other text-to-image models like FLUX or Z-Image?
Lens achieves comparable quality to larger models through training data quality rather than parameter count. It was trained on Lens-800M, a dataset of 800 million image-text pairs with dense captions averaging 109 words each, generated by GPT-4.1. This gives the model more information per training example than models trained on short alt-text captions, reducing the compute needed to reach the same output quality.

What text encoder does Lens use?
Lens uses GPT-OSS 20B, a text encoder that concatenates features from multiple internal layers (layers 4, 12, 18, and 24) to produce richer text representations. This multi-layer approach strengthens prompt adherence and enables multilingual generalization even though the model was trained on English captions only.

What resolution and aspect ratios does Lens Turbo support?
Lens supports resolutions up to 1440x1440 and nine standard aspect ratios from 1:2 to 2:1 natively. This workflow defaults to 1280x720 (16:9 landscape). Change the latent dimensions to target a different ratio or resolution.

Does Lens Turbo support multilingual prompts?
Yes. The GPT-OSS text encoder generalizes across multiple languages. You can write prompts in languages other than English and the model produces images with consistent quality, even though the training corpus was English-only.

Is Microsoft Lens Turbo licensed for commercial use?
Lens weights are released on Hugging Face under the MIT license. However, Microsoft states the model is released for research purposes and is not intended for product or service deployment. Review the full license and usage terms on the Hugging Face model page before using outputs in commercial projects.

How to run Microsoft Lens Turbo online?
You can run Microsoft Lens Turbo online through Floyo. No installation, no setup, no local GPU needed. Open the workflow in your browser, write your prompt, and hit run. Free to try.

WHY FLOYO?

Floyo is the only platform with team collaboration for ComfyUI in the browser. You run workflows with no install. You share run history, assets, and models across your team. You pay only when you generate. Floyo supports open-source and closed-source models.

A designer runs an edit and likes the result. A teammate opens that exact run from shared history and keeps going. No file handoffs. No version confusion.

For studios and enterprise teams, Floyo adds private workspaces, pooled resources, and a team usage dashboard. Other ComfyUI cloud tools run for one person at a time. Floyo runs for the whole team, with transparent per-generation costs.

Ready to try it?
Write a detailed prompt and hit run.

→ Launch Workflow, Free

Questions? Watch the free course or check the FAQ above.