Echo-4o-Image Dataset

Paper | Project Page | Code

Introduction

Echo-4o-Image is a 180K-scale synthetic dataset generated by GPT-4o, designed to advance open-source models in image generation. While real-world image datasets are valuable, synthetic images offer crucial advantages, especially in addressing blind spots in real-world coverage:

• Complementing Rare Scenarios: Synthetic data can generate examples for scenarios less represented in real-world datasets, such as surreal fantasy or multi-reference image generation, which are common in user queries.
• Clean and Controllable Supervision: Unlike real-world data, which often contains complex background noise and misalignment between text and image, synthetic images provide pure backgrounds and long-tailed supervision signals, facilitating more accurate text-to-image alignment.

This dataset was instrumental in fine-tuning the unified multimodal generation baseline Bagel to obtain Echo-4o, demonstrating strong performance across standard benchmarks. Furthermore, Echo-4o-Image consistently enhances other foundation models (e.g., OmniGen2, BLIP3-o), highlighting its strong transferability.

Echo-4o-Image Dataset Details

Echo-4o-Image is a large-scale synthetic dataset distilled from GPT-4o, containing approximately 179,000 samples. It spans three distinct task types:

• 38K surreal fantasy generation tasks: Designed to address imaginative content.
• 73K multi-reference image generation tasks: For scenarios requiring multiple visual cues.
• 68K complex instruction execution tasks: To improve adherence to detailed textual prompts.

For better visualization, an online gallery showcasing representative samples from our dataset is available: Online Gallery

Data Structure

The dataset typically organizes data within compressed packages (e.g., .tar.gz files referenced in configs). Inside these packages, data is arranged as follows:

- package_idx/
--- package_idx.json # metadata for samples in this package
--- images/
----- 00001.png
----- 00002.png
...

Usage

This dataset can be used to train and fine-tune text-to-image models, extending capabilities to support multi-reference datasets.

Training

The training process extends existing frameworks (e.g., Bagel's capabilities).

1. Data Preparation: Follow data preparation guidelines, ensuring multi-reference data adheres to the expected format.
2. Training Process: Training scripts use interfaces and parameters similar to established models (e.g., Bagel), allowing for seamless integration with existing training commands and configurations.

Inference

• Text-to-Image Tasks: For standard text-to-image generation, follow the inference process of base models (e.g., Bagel).
• Multi-Reference Tasks: Specific examples and guides for tasks involving multiple references are provided in the official GitHub repository.

Code and Supporting Files

The associated GitHub repository provides crucial supporting files for working with the dataset:

• Attributes and Subjects: ./code/attributes_and_subjects.json contains dictionaries defining various attributes and subjects used in the dataset.
• Range-sensitive filtering: ./code/range_sensitive_filter.json contains metadata for data filtering, and ./code/data_filter.py converts it for use in dataloaders.
• Data Loader: ./code/dataloader.py provides an example of how to load the data into image pairs, incorporating filtering and balanced resampling.

Evaluation Benchmarks

The paper introduces two novel benchmarks for rigorously evaluating image generation capabilities:

• GenEval++: Increases instruction complexity and uses an automated evaluator (powered by GPT-4.1) to mitigate score saturation and provide a more accurate assessment of text-to-image instruction following.
• Imagine-Bench: Focuses on imaginative content, offering a comprehensive evaluation of conceptual creativity and visual consistency across dimensions like fantasy fulfillment, identity preservation, and aesthetic quality.

Detailed guides for these benchmarks can be found in the EVAL section of the GitHub repository.

Acknowledgements

We would like to thank the following open-source projects and research works:

• Bagel
• BLIP3o
• OmniGen2

Citation

If you find this dataset or the associated work useful for your research, please cite the paper:

@article{ye2025echo4o,
      title={Echo-4o: Harnessing the Power of GPT-4o Synthetic Images for Improved Image Generation}, 
      author={Junyan Ye, Dongzhi Jiang, Zihao Wang, Leqi Zhu, Zhenghao Hu, Zilong Huang, Jun He, Zhiyuan Yan, Jinghua Yu, Hongsheng Li, Conghui He, Weijia Li},
      journal={https://arxiv.org/abs/2508.09987},
      year={2025},
}