Face Description

Says who is in the picture in words: one description per detected face.

The face description node produces a natural-language description of the faces detected in an asset, using a vision-language model. Its detections port carries a sequence — one element per detected face — and the node describes each element in turn, so a picture with three faces comes back with three descriptions in the same order. A sequence with no elements means there is nothing to describe and no model call is made.

Kind

facedescription

Applies to

Image (video is stubbed / not yet implemented)

Input ports

detections (many) — connect the detections port of Face Detection; the crops travel with the elements, so the picture itself is not wired in

Output ports

descriptions (many — one per incoming face, aligned with the input sequence)

Requirements

A vision-language / LLM model service reachable from the worker. CPU on the worker; the model runs in the service.

Persists to

asset_json_comp + asset_node_result ledger

Configuration

The Face Description settings panel: Max Face Angle (deg)
Figure 1. The node’s settings in the pipeline editor

Uses the model-service connection shared with the LLM/captioning family. There is nothing to configure about where the faces come from — connect the detections port and the node follows it.

Seeing it run

Turn on Debug Mode and every node keeps what it produced, on the card itself. Below is a real run over a frame with two people in it, fed the boxes Facedetect found in that same frame.

The facedescription node after a run

Two boxes in, two descriptions out, aligned by element index — that alignment is why a face that cannot be described still emits an empty document rather than being skipped. Each description is the JSON the prompt asks for: eye state, mouth state, an age estimate, whether the face is frontal or in profile.

Use Cases

  • Searchable people descriptions — "man with glasses, outdoors" style text for retrieval.

  • Accessibility / alt text for images containing people.

  • Review context — a quick textual summary of who/what is in a flagged image.

Looking for something else?