Task and Config Matrix

This page is not an exhaustive metric catalog. Its purpose is to map tasks to starter configs so you can begin from an existing setup instead of authoring YAML from scratch.

Object-Centric

Starter Config	Typical task	Notes
object_centric/subject_add.yaml	subject_add	Add a subject or object
object_centric/subject_remove.yaml	subject_remove	Remove a target subject
object_centric/subject_replace.yaml	subject_replace	Replace a subject
object_centric/color_alter.yaml	color_alter	Modify color attributes
object_centric/material_alter.yaml	material_alter	Modify material attributes
object_centric/size_adjustment.yaml	size_adjustment	Resize a target
object_centric/text_editing.yaml	text_editing	Edit text inside an image
object_centric/cref.yaml	cref	Character or subject reference tasks
object_centric/oref.yaml	oref	Object reference tasks

Human-Centric

Starter Config	Typical task	Notes
human_centric/ps_human.yaml	ps_human	Portrait or human-appearance editing
human_centric/motion_change.yaml	motion_change	Human pose or motion editing

VLM-as-a-Judge

Starter Config	Use case	Notes
vlm_as_a_judge/openai.yaml	API backend	OpenAI-style chat endpoint
vlm_as_a_judge/gemini.yaml	Google-style backend	Gemini-compatible backend
vlm_as_a_judge/vllm.yaml	Self-hosted backend	Local or internal vLLM service
vlm_as_a_judge/eval_if.yaml	Benchmark eval	Instruction-following judge prompt
vlm_as_a_judge/eval_vc.yaml	Benchmark eval	Visual-consistency judge prompt
vlm_as_a_judge/eval_vq.yaml	Benchmark eval	Visual-quality judge prompt

Benchmark Keys

eval reads benchmark definitions from configs/datasets/bmk.json. Repository-visible keys include:

• openedit
• vc_reward
• editscore_consistency
• editscore_prompt_following
• editreward_visual_quality

How to use this matrix

1. Decide whether you need structured scores or pairwise winners.
2. Choose the pipeline family.
3. Start from the closest existing config instead of creating a blank YAML file.