Microsoft Lens in ComfyUI: setup and Z Image comparison

This tutorial builds a ComfyUI workflow for Microsoft Lens, a 3.8-billion-parameter text-to-image model. The presenter loads the diffusion model, selects the Lens type for the GPT OSS NVFP4 text encoder, and uses the Flux 2 VAE. The walkthrough covers generating images on your own hardware, with a smaller MX FP8 Turbo variant shown for cards with less VRAM.

Lens Turbo uses four sampling steps in the demonstration. The presenter reports roughly five seconds per image and prefers its results to the default model, which takes about four times longer in that test. These timings describe the demonstrated setup rather than a general performance guarantee.

The workflow also covers image-to-image generation, masked inpainting, and prompt-based masks with SAM 3. Image and latent upscaling receive separate examples, followed by tiled Ultimate SD Upscale and Seed VR 2 comparisons. The presenter finds that second passes add detail, while Seed VR 2 changes the image less.

The stated resolution limit is 1440×1440. The presenter also tries 2048×2048, but warns that pushing further produces duplicates and other artifacts. Spelling remains inconsistent. Comparisons with Z Image Turbo favor Lens for prompt adherence in the selected examples; those judgments reflect the presenter's prompts and preferences.