Skip to main content

How to Choose a Pipeline

If you remember only one rule, remember this:

  • Use human-centric or object-centric when you want structured scores for a single edited image
  • Use vlm-as-a-judge when you want to compare two candidate edits directly

Quick selection matrix​

QuestionRecommended pipelineOutput
Did a single edit satisfy an object-level instruction?object-centricscope -> metric -> score
Did a single human edit preserve identity, appearance, or local consistency?human-centricscope -> metric -> score
Which of two candidate edits is better?vlm-as-a-judgewinner + raw_responses

When to choose object-centric​

Typical tasks include:

  • subject_add
  • subject_remove
  • subject_replace
  • color_alter
  • material_alter
  • size_adjustment
  • text_editing
  • cref / oref

Key characteristics:

  • Scores a single edited image at a time
  • Relies primarily on parser-grounder plus general-purpose visual metric pipes
  • Starter configs live in configs/pipelines/object_centric/

When to choose human-centric​

Choose this family for:

  • ps_human
  • motion_change
  • other tasks that depend on human-local consistency

Key characteristics:

  • Loads face-detector, human-segmenter, and hair-segmenter
  • Separates edit_area from unedit_area
  • Uses attribute-aware measurement rubrics to decide which metrics actually run

When to choose vlm-as-a-judge​

Choose this family for:

  • benchmark pairwise evaluation
  • direct A/B comparison of candidate outputs
  • direct construction of pairwise training pairs

Key characteristics:

  • Expects instruction + input_image + edited_images
  • Returns Image A, Image B, Tie, or Failed
  • Starter configs live in configs/pipelines/vlm_as_a_judge/

A common confusion​

Do not confuse "I want train pairs" with "which pipeline should I choose."

  • train-pairs is a post-processing command
  • The earlier choice between annotation and eval depends on whether you want structured scores or pairwise winners first

Continue with: