Skip to main content

Human-Centric Pipeline

human-centric is the pipeline family for person and body editing. Compared with object-centric, the difference is not just a different metric list. The entire runtime also depends on human-specific experts and cropped local regions.

What it is designed to measure​

Typical questions include:

  • Was facial identity preserved?
  • Did facial geometry or facial texture degrade?
  • Is the edited hair region consistent?
  • Did body pose and appearance remain coherent?
  • Which unedited regions still need evaluation after a large human edit?

Starter configs​

  • configs/pipelines/human_centric/ps_human.yaml
  • configs/pipelines/human_centric/motion_change.yaml

What makes this family different​

Unlike object-centric, this family requires expert_configs for:

  • face-detector
  • human-segmenter
  • hair-segmenter

It also separates runtime measurements into:

  • edit_area
  • unedit_area

The pipeline uses parsed edit attributes to assemble a measurement rubric, which determines which metrics should actually run for a given sample.

Minimal example​

cd <PROJECT_ROOT>
autopipeline annotation \
  --edit-task motion_change \
  --pipeline-config-path <PROJECT_ROOT>/configs/pipelines/human_centric/motion_change.yaml \
  --user-config <PROJECT_ROOT>/configs/pipelines/user_config.yaml \
  --save-path <PROJECT_ROOT>/data/c_annotated_group_data

What you will see in the results​

The output is still grouped JSONL, but the score structure usually includes both:

  • edit_area
  • unedit_area

Some metrics may still be null. That does not necessarily indicate a broken run. Common reasons include:

  • face bounding boxes were not detected
  • segmentation masks were unavailable
  • the measurement rubric skipped a metric for that edit type

When to prefer it​

If your quality criteria depend on human-local details, use human-centric first instead of forcing the task into object-centric.

If your real question is which of two candidate images is better rather than how to score one image structurally, continue with VLM-as-a-Judge.