Skip to main content

Object-Centric Pipeline

object-centric is the easiest pipeline family to start with. It is designed for object-level evaluation of a single edited image and focuses on:

  • whether the edit target can be parsed and localized correctly
  • whether the unedited region remains stable
  • whether structure and semantics remain plausible for the task

Supported starter configs​

The repository already includes starter configs such as:

  • subject_add.yaml
  • subject_remove.yaml
  • subject_replace.yaml
  • color_alter.yaml
  • material_alter.yaml
  • size_adjustment.yaml
  • text_editing.yaml
  • cref.yaml
  • oref.yaml

What you need before running it​

  • an object-centric pipeline YAML
  • a valid user_config.yaml
  • the candidate pool JSON for the task

annotation resolves the candidate pool automatically from --edit-task, so the task name and the chosen config must match.

Configuration shape​

The two most important config blocks are:

  • parser_grounder_config
  • metric_configs

In other words, this family usually does not require expert_configs.

Minimal example​

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

What it actually does​

  1. Normalizes instruction, input_image, and edited_images[0] into a common input schema
  2. Uses parser-grounder to parse the instruction and localize relevant regions
  3. Resolves each metric to a registered pipe from metric_configs
  4. Computes scores for edit_area, unedit_area, or both, depending on the config
  5. Aggregates results into grouped JSONL

Good fit vs poor fit​

Good fit:

  • structured scoring of a single edited image
  • object or scene-level edits
  • metric-driven filtering and data construction

Poor fit:

  • tasks dominated by facial identity or body-local consistency
  • problems that are fundamentally pairwise comparison tasks

If your problem is primarily about human consistency, continue with Human-Centric.