HF2026Repo/INTENT-Edit-Bench

Dataset Summary

INTENT-Edit-Bench is a multimodal benchmark for instruction-following image editing with object-level edit plans. Each example pairs an input image (primary_image) with an editing instruction (instruction) and a structured plan (prepare_edit_plan) describing compulsory/optional objects and constraints.

• Hugging Face dataset: https://huggingface.co/datasets/HF2026Repo/INTENT-Edit-Bench
• License: CC BY 4.0 (https://choosealicense.com/licenses/cc-by-4.0/)
• Version: 1.0
• Size / split: 1,024 examples, train only

Dataset Structure

Data Files

• Train parquet
  • Data/train/INTENT_Edit.parquet (1,024 rows)
  • Contains all text fields and one image field (primary_image)
  • Encodes primary_image in HF image struct format (bytes + path) for viewer compatibility
• Utility script in this directory
  • expand_parquet.py (expands parquet back into CSV + JPG tree)

primary_image stores an image object (embedded JPG bytes + relative path) compatible with HF viewer. Paths follow the on-disk layout: Primary_Image/<paradigm>/<domain>/<full_id>.jpg.

Splits and Sizes

• Split: train only
• Total rows: 1024
• Category counts:
  • SocialNorms: 125
  • ObjectDynamics: 125
  • TemporalEvolution: 113
  • CulturalSymbols: 124
  • CommonSense: 125
  • DataAnalysis: 145
  • RuleConstraints: 127
  • ScientificPrinciples: 140

Data Fields (Parquet schema)

Data/train/INTENT_Edit.parquet contains:

• full_id (string): Unique example id (e.g. CommonSense_0, CulturalSymbols_0).
• domain (string): Domain label (e.g. CommonSense, CulturalSymbols).
• paradigm (string): Higher-level grouping (e.g. Formal_Reasoning, Real-world_Knowledge).
• instruction (string): Natural-language edit instruction.
• prepare_edit_plan (stringified JSON): Object-level plan items with fields like object_id, object_name, status (COMPULSORY/OPTIONAL), and intention.
• prepare_object_list (stringified JSON): Object name list aligned with the plan.
• primary_image (image): HF-compatible image object containing JPG bytes and relative path metadata.

Intended Use

• Benchmarking: Evaluate instruction-following image editing and planning under object-level constraints.

Supported Tasks

• Image editing / image-to-image generation: Given primary_image + instruction, generate an edited image that satisfies the intent and constraints.
• Edit planning / structured generation: Given primary_image + instruction, generate a plan similar to prepare_edit_plan.

Quick Start

If you are using the HF dataset, prefer loading from the Hugging Face hub:

from datasets import load_dataset

ds = load_dataset("HF2026Repo/INTENT-Edit-Bench", split="train")
print(ds.features)

To expand parquet back to CSV + images:

cd INTENT-Edit-Bench
python expand_parquet.py

Expand Output Structure

After running expand_parquet.py, the dataset is restored to the following structure:

INTENT-Edit-Bench/
├── INTENT-Edit-Bench.csv
└── Primary_Image/
    ├── Formal_Reasoning/
    │   ├── CommonSense/*.jpg
    │   ├── DataAnalysis/*.jpg
    │   ├── RuleConstraints/*.jpg
    │   └── ScientificPrinciples/*.jpg
    └── Real-world_Knowledge/
        ├── CulturalSymbols/*.jpg
        ├── ObjectDynamics/*.jpg
        ├── SocialNorms/*.jpg
        └── TemporalEvolution/*.jpg

Optional arguments:

• --parquet: input parquet path (default Data/train/INTENT_Edit.parquet)
• --output-dir: output root directory (default current directory)
• --csv-output: output CSV filename (default INTENT-Edit-Bench.csv)
• --image-column: image column name in parquet (default primary_image)
• --limit: export only first N rows (debug/testing)