Sapiens2 Seg

Sapiens2 Seg is a body-part segmentation workflow from Meta's Sapiens2 family. It predicts a class label for each pixel in a human-centric image, so you can separate regions like apparel, hair, face and neck, arms, legs, hands, feet, shoes, lips, teeth, tongue, and background.

The practical value is the mask, not the architecture. In the browser demo, you can upload an image, choose a model size, inspect the annotated overlay, and download the raw label map as an NPY file. That makes it easier to use the result as a preprocessing step for image editing, dataset cleanup, try-on experiments, avatar work, or any workflow that needs more detail than "person versus background."

Start with the Hugging Face Space and upload a clear full-body or upper-body photo. Choose the 1B model first, then check whether the overlay separates clothing, hands, hair, and face regions in a way that would actually help your downstream task.

If the browser result is useful, move to the model page and Sapiens2 GitHub repo for the local path. The local workflow downloads the checkpoint with the Hugging Face CLI, then runs the segmentation demo script after you set the input folder, output folder, and model size.

Expect CUDA/GPU setup rather than a casual CPU run. Also treat this as a body-part masking tool, not an identity or biometric analysis tool; the Sapiens2 license places restrictions around sensitive and biometric uses.

• Create body-part masks before targeted image editing or inpainting.
• Separate clothing, hair, skin, and limb regions for fashion or catalog workflows.
• Add a human-centric preprocessing step before normal, albedo, pointmap, or avatar pipelines.
• Bootstrap labels for datasets where manual body-part annotation would be slow.
• Check whether a people-image pipeline fails on hands, occlusion, clothing, or small body regions.

Try the demo

View model page

Neural Expedition · Useful open-source AI, curated without hype.