Microsoft's Lens Beats 6B Models at 3.8B and Now Runs in ComfyUI
Comfy-Org repackages Microsoft's MIT-licensed Lens model into drop-in ComfyUI files, with native PR integration pending merge
Takeaways
- Comfy-Org repackaged Microsoft's MIT-licensed Lens model into drop-in ComfyUI
.safetensorsfiles, with no access request required. - Lens is a 3.8B text-to-image model that matches or beats 6B+ models while using only ~19.3% of Z-Image's training compute.
- Three variants available: Lens-Base (50 steps), Lens (20 steps, RL-tuned), and Lens-Turbo (4 steps distilled).
- A native integration PR (#14077) is open against the main ComfyUI repo; early adopters can run it today via a manual git checkout.
- Community tests rate it strong for nature/animals (8/10) but weak for human subjects (4/10); no formal benchmarks yet.
- Released for research only -- not intended for product or service deployment per Microsoft's own license terms.
Microsoft's Lens is a research model that quietly punches well above its weight class. Now, Comfy-Org has repackaged it into ComfyUI-ready files, and a pull request for native integration is open against the main ComfyUI repo. The model itself was already MIT-licensed and publicly available from Microsoft. What changed is that using it in ComfyUI just got a lot easier.
A 3.8B model that competes with 6B+ models
Lens is a 3.8B-parameter text-to-image model that achieves performance competitive with, and in several cases surpassing, state-of-the-art models with more than 6B parameters across various benchmarks, while requiring significantly less training compute. The research angle is squarely about efficiency: Lens requires only about 19.3% of the training compute used by Z-Image.
To put that in concrete terms: Z-Image requires approximately 314K H800 GPU hours for pre-training. Lens is trained with 128 A100 GPUs, requiring only about 19.3% of the training cost of Z-Image. That's a meaningful gap, and it comes from rethinking the training data rather than the architecture.
How it achieves that efficiency
Microsoft maximized data information density per training batch by training on Lens-800M, a dataset of 800M densely captioned image-text pairs whose captions are generated by GPT-4.1 and contain approximately 109 words on average, providing richer semantic supervision than conventional short captions. Each batch is also constructed from images at multiple resolutions and aspect ratios, widening the effective visual coverage of each optimization step.