ImageAgent Styling Dataset

Dataset Summary

The ImageAgent Styling Dataset is a large-scale synthetic collection of image transformation trajectories designed for training and evaluating vision-language models on compositional image editing tasks. The dataset contains 30,000 samples across three complexity variants, each comprising original images, edited images, structured action plans with chain-of-thought reasoning, image analysis, and multi-dimensional reward scores.

The dataset was created to support research in agentic planning with offline reinforcement learning for image styling, as described in the associated paper.

Dataset Variants

Naming note: In the associated paper, the Normal (Regular) split — called Regular in the paper. The two names refer to the same dataset. The directory and action library files use normal; the paper text uses Regular for clarity within its complexity hierarchy (Simple → Regular → Complex).

---

Dataset Structure

image-agent-styling/
├── actions/
│   ├── action_library_simple.json    # 10 atomic actions (Simple dataset)
│   ├── action_library_normal.json    # 20 actions incl. constraints (Normal/Regular)
│   └── action_library_complex.json   # 30 actions, integrated constraints (Complex)
├── simple/                           # 10K samples, 5 parquet shards (~18 GB)
├── normal/                           # 10K samples, 5 parquet shards (~15 GB)
└── complex/                          # 10K samples, 5 parquet shards (~16 GB)

---

Data Fields

Each sample contains:

---

Data Generation Pipeline

Each trajectory is produced by a 5-stage pipeline:

1. Base Image Generation: HiDream-I1-Dev generates source images from theme-specific prompts across diverse visual categories (landscapes, portraits, cityscapes, fantasy, etc.).
2. Context Extraction: GPT-4o analyzes each image to extract structured visual context (objects, colors, style, mood, spatial relationships).
3. Action Plan Generation: GPT-4o generates a structured action plan with per-step chain-of-thought reasoning, drawn from the dataset-specific action library.
4. Image Editing: Qwen-Image-Edit (20B) applies the action plan to produce the edited image.
5. Reward Evaluation: GPT-4o scores each trajectory on four dimensions (action plan quality, edit quality, reasoning quality, overall) on a 0–5 scale.

Human evaluation (3 independent annotators, 3,000 samples) validated 77% Pass rate and 14.9% Partial across all three splits. See the associated paper for full human evaluation methodology and inter-annotator agreement statistics.

---

Usage

Loading the Dataset

from datasets import load_dataset

# Load Simple split (atomic 1-2 step edits)
simple = load_dataset("subhojyoti1990/image-agent-styling", data_dir="simple", split="train")

# Load Normal (Regular) split (compositional 3-5 step, 10 interior themes)
# Note: called "Regular" in the paper
normal = load_dataset("subhojyoti1990/image-agent-styling", data_dir="normal", split="train")

# Load Complex split (advanced, 83 diverse themes)
complex_ds = load_dataset("subhojyoti1990/image-agent-styling", data_dir="complex", split="train")

Accessing Action Libraries

from huggingface_hub import hf_hub_download
import json

for split in ["simple", "normal", "complex"]:
    path = hf_hub_download(
        repo_id="subhojyoti1990/image-agent-styling",
        filename=f"actions/action_library_{split}.json",
        repo_type="dataset"
    )
    with open(path) as f:
        lib = json.load(f)
    print(f"{split}: {len(lib['actions'])} actions defined")

Processing a Sample

import json

sample = normal[0]

# Images are PIL Image objects
original = sample['original_image']
edited   = sample['edited_image']

# JSON fields need parsing
action_plan   = json.loads(sample['action_plan'])
analysis      = json.loads(sample['analysis'])
reward_scores = json.loads(sample['reward_scores'])
prompt_meta   = json.loads(sample['prompt'])

# Text fields
print(sample['instruction'])   # Natural language instruction
print(sample['edit_prompt'])   # Full editing prompt

# Reward (0-5 scale)
overall_score = reward_scores['scores']['overall']['score']

---

Dataset Statistics

---

Supported Tasks

• Image-to-Image Transformation: Train/evaluate models on natural language–driven image editing
• Agentic Action Planning: Learn to generate structured multi-step tool-use plans
• Chain-of-Thought Reasoning: Supervisory signal for per-step reasoning in visual tasks
• Offline RL / Reward-Weighted Fine-Tuning: Reward scores enable reward-weighted (RW/SW) and DPO training directly from the dataset

---

Responsible AI Metadata

Synthetic Data

All data in this dataset is fully synthetic. There are no real photographs, real people, or personally identifiable information. Source images are generated by HiDream-I1-Dev; edited images are produced by Qwen-Image-Edit. Editing instructions and action plans are generated by GPT-4o.

Data Provenance

• Source seeds: Theme libraries and category taxonomies (available in actions/); no external datasets were used as image sources
• Image generation: HiDream-I1-Dev (AI-generated base images)
• Planning & scoring: GPT-4o (gpt-4o-2024-05-13) via Azure OpenAI API
• Image editing: Qwen-Image-Edit (20B instruction-following VLM)
• Human validation: 3 independent annotators with image quality experience evaluated 1,000 samples per split (3,000 total) using a Pass/Partial/Fail rubric on 4 dimensions (edit quality, action plan quality, reasoning quality, overall). Inter-annotator agreement: 62–66% exact, <11% complete disagreement.

Data Limitations

• English-only: All instructions, action plans, and reasoning chains are in English
• Synthetic distribution: Images are AI-generated; distribution may differ from real-world photographs in texture, lighting physics, and object proportions
• Theme coverage: Simple (31 location types), Normal (Regular) (10 interior design styles), Complex (83 themes) — does not cover all visual domains
• GPT-4o scoring bias: Reward scores are assigned by GPT-4o, which introduces model-specific evaluation biases. Human studies show moderate top-2 accuracy (78.3%) between GPT-4o rankings and human preferences, with weak exact-winner agreement — scores should be treated as noisy proxies, not ground truth
• HiDream model biases: Generated images may reflect biases present in HiDream's training data (over-representation of certain aesthetics, Western visual norms)
• No temporal/video data: Static images only; not suitable for video editing tasks

Data Biases

• Visual aesthetic bias: Interior design themes (Normal (Regular) split) skew toward contemporary Western interior styles
• Image quality distribution: Stratified sampling for human evaluation covered high (reward ≥ 4.0, 40%), medium (3.0–4.0, 40%), and low (<3.0, 20%) quality tiers; the full dataset includes noisy/low-quality trajectories by design for reward-weighted training
• Teacher model bias: GPT-4o's action planning preferences are encoded into the training signal; models fine-tuned on this data may inherit GPT-4o's compositional planning style

Personal and Sensitive Information

None. All images are fully AI-generated. No real people, faces, biometric data, locations, or PII are present in this dataset.

Intended Use Cases

Valid uses (validity established):

• Fine-tuning vision-language models for image editing and styling
• Offline RL / reward-weighted training for compositional planning
• Research on agentic reasoning in visual tasks
• Benchmarking action planning and tool-use in VLMs

Not recommended for:

• High-stakes decisions (medical, legal, safety-critical)
• Face recognition or person identification
• Training models intended to deceive or manipulate (deepfakes, misinformation)
• Direct deployment in consumer applications without further safety evaluation

Social Impact

Positive: Enables research into interpretable, structured AI image editing; provides a reproducible benchmark with explicit reasoning chains; supports open-source model development as an alternative to proprietary systems.

Negative / Risks: Models trained on this dataset could be used to generate misleading or manipulated imagery at scale. The reward signal from GPT-4o may encode proprietary aesthetic preferences that are hard to audit. All use should comply with applicable AI ethics guidelines and the MIT license terms.

---

Paper

Agentic Planning with Reasoning for Image Styling via Offline RL
Subhojyoti Mukherjee, Stefano Petrangeli, Branislav Kveton, Trung Bui, Franck Dernoncourt, Arko Mukherjee
Adobe Research
arXiv: https://arxiv.org/abs/2603.07148

---

Citation

@misc{mukherjee2026agenticplanningreasoningimage,
      title={Agentic Planning with Reasoning for Image Styling via Offline RL}, 
      author={Subhojyoti Mukherjee and Stefano Petrangeli and Branislav Kveton and Trung Bui and Franck Dernoncourt and Arko Mukherjee},
      year={2026},
      eprint={2603.07148},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2603.07148}
}

---

Contact

For questions or issues: subhomuk@adobe.com