Tencent Cloud Portrait Segmentation

Problem

In image workflows, a common task is to separate a person from a busy background for poster composites, video assets, e-commerce product images, or virtual character work. Manual masking or asking a general model to “look” at the image can produce unstable results and cannot reliably prove that the output is a real segmentation. This skill reduces the task to one synchronous API call: pass in an image, detect the full human outline, and return a portrait segmentation result, avoiding the risk of mistaking a plausible-looking result for the final cutout.

How It Works

The skill is built on Tencent Cloud's SegmentPortraitPic capability and runs through the core script scripts/main.py. It accepts an input image, calls the Tencent Cloud portrait segmentation service, performs human outline detection and cutout processing, and returns the result synchronously. It requires the portrait segmentation service to be enabled, an API key to be configured, environment variables to be set, and a runtime compatible with Python 3.7+ and tencentcloud-sdk-python; the provided notes say the script can install dependencies automatically. The execution contract for an agent is also explicit: when a user supplies an image and intends portrait segmentation, the skill should run the script directly, not merely print the path, not ask for confirmation, and not manually install dependencies; if the call fails, it must return a clear error and must not let the model guess or fabricate segmentation content.

Boundaries

This skill is suited for portrait segmentation on one or a few images, with an input size limit of ≤5MB and a synchronous return model. It is not a general-purpose segmentation tool, and it does not promise support for arbitrary poses, multiple overlapping people, heavy occlusion, or non-person subjects. In production use, check service enablement, key permissions, and network reachability; preserve original error messages for failed calls so the issue can be traced to parameters, quotas, service state, or the input image. It fits into an existing media pipeline as a deterministic step from raw image to portrait asset.