Source-linked AI summary
GPT-IMAGE-EDIT-1.5M: A Million-Scale, GPT-Generated Image Dataset
Yuhan Wang, Siwei Yang, Bingchen Zhao, Letian Zhang, Qing Liu, Yuyin Zhou, Cihang Xie
TL;DR
Open-source image editing lacks sufficiently large, diverse, and high-quality data, while leading proprietary models remain closed. The paper introduces GPT-IMAGE-EDIT-1.5M by refining OmniEdit, HQ-Edit, and UltraEdit with GPT-4o, and fine-tunes open-source models on it. The resulting FluxKontext reaches state-of-the-art open-source performance across multiple benchmarks, including 7.236 on GEdit-EN and 3.80 on ImgEdit-Full.
Problem
Open-source image editing lacks large, diverse, high-quality training data, while leading proprietary models and their datasets are closed.
Method
GPT-IMAGE-EDIT-1.5M unifies OmniEdit, HQ-Edit, and UltraEdit by regenerating images and selectively rewriting instructions with GPT-4o.
Results
7.236 on GEdit-EN and 3.80 on ImgEdit-Full establish strong open-source benchmark performance for FluxKontext trained on GPT-IMAGE-EDIT-1.5M.
Takeaways & Limitations
The released dataset and models provide a public resource for further research in instruction-guided image editing.
Abstract
from arXiv · showhide
Recent advancements in large multimodal models like GPT-4o have set a new standard for high-fidelity, instruction-guided image editing. However, the proprietary nature of these models and their training data creates a significant barrier for open-source research. To bridge this gap, we introduce GPT-IMAGE-EDIT-1.5M, a publicly available, large-scale image-editing corpus containing more than 1.5 million high-quality triplets (instruction, source image, edited image). We systematically construct this dataset by leveraging the versatile capabilities of GPT-4o to unify and refine three popular image-editing datasets: OmniEdit, HQ-Edit, and UltraEdit. Specifically, our methodology involves 1) regenerating output images to enhance visual quality and instruction alignment, and 2) selectively rewriting prompts to improve semantic clarity. To validate the efficacy of our dataset, we fine-tune advanced open-source models on GPT-IMAGE-EDIT-1.5M. The empirical results are exciting, e.g., the fine-tuned FluxKontext achieves highly competitive performance across a comprehensive suite of benchmarks, including 7.24 on GEdit-EN, 3.80 on ImgEdit-Full, and 8.78 on Complex-Edit, showing stronger instruction following and higher perceptual quality while maintaining identity. These scores markedly exceed all previously published open-source methods and substantially narrow the gap to leading proprietary models. We hope the full release of GPT-IMAGE-EDIT-1.5M can help to catalyze further open research in instruction-guided image editing.
1 Introduction
GPT-IMAGE-EDIT-1.5M addresses the shortage of large, diverse, high-quality open-source image-editing data by refining three existing datasets with GPT-4o. Fine-tuned models achieve strong benchmark results, including 7.24 on GEdit-EN-full and 3.80 on ImgEdit.
- Closed-source models and training data limit broader research despite their strong semantic understanding and photorealistic editing.
- Large public datasets still contain noisy instructions, misaligned image pairs, and limited editing diversity, hindering open-source model development.
- Over 1.5 million samples unify OmniEdit, HQ-Edit, and UltraEdit through GPT-4o-based image regeneration and instruction refinement.
- 3.40 on OmniEdit follows instruction regeneration, improving on the 3.24 score from output-image regeneration alone.
- 7.236 on GEdit-EN and 3.80 on ImgEdit-Full place the fine-tuned FluxKontext among the strongest open-source models.
2 Related Works
Prior work established instruction-guided editing, data-centric curation, evolving generative architectures, multimodal instruction encoders, and complementary evaluation benchmarks. GPT-IMAGE-EDIT-1.5M extends the data-centric direction by systematically refining existing datasets with GPT-4o.
- Instruction-Guided Image Editing: InstructPix2Pix framed instruction-guided image editing as supervised learning using synthetic instruction, source-image, and edited-image triplets.
- Data-Centric Advancements: HQ-Edit and ShareGPT-4o-Image demonstrate data-centric strategies that use proprietary models to create higher-quality or frontier-model-distilled editing data.
- Data-Centric Advancements: GPT-IMAGE-EDIT-1.5M uses GPT-4o for systematic refinement and enhancement of existing large-scale datasets, rather than generation alone.
- Architectural Evolution: From Diffusion to Flow Matching: Flow-matching models learn a continuous velocity field and provide the architectural basis for FLUX.1 Kontext’s unified generation-and-editing approach.
- Enhancing Semantic Control with MLLM Encoders: MLLM encoders address limitations of standard CLIP encoders in parsing complex spatial, relational, and compositional instructions.
- Evaluation Benchmarks: The evaluation suite spans GEdit-Bench-EN, ImgEdit, Complex-Edit, and OmniContext to assess diverse editing capabilities and quality dimensions.
3 Data Curation
The curation pipeline merges three datasets, standardizes their metadata, regenerates images and instructions, and filters low-quality outputs. Dataset-specific processes include rewriting about 10% of OmniEdit instructions and regenerating about 50% of HQ-Edit inputs.
- The pipeline replaces operation-specific expert collection with a minimalist approach using a preexisting general-purpose image-editing model.
- OmniEdit, HQ-Edit, and UltraEdit are merged into a consistent corpus, with gpt-image-1 regenerating or augmenting image-instruction pairs.
- Fixed 1:1, 3:2, and 2:3 aspect ratios, followed by automatic artifact and padding filtering, standardize generated outputs.
- Instruction Re-writing: Approximately 10% of OmniEdit instructions were rewritten with GPT-4o using the original input and updated output images.
- Input Image Regeneration: Approximately 50% of HQ-Edit input images were regenerated with GPT-Image-1, with corresponding outputs produced from those inputs.
- Complex-Edit style instruction: Complex-Edit-style instructions were generated using about 50% of OmniEdit input images, while very complex prompts could reduce realistic appearance.
4 Experiments
Experiments evaluate the enhanced FluxKontext model across general, compositional, and context-aware editing benchmarks, alongside ablations of data curation, Complex-Edit data, and text encoders. The curated data and refinement strategies yield strong benchmark performance, while raw complex instructions can harm identity preservation.
- Experimental Setup: The primary model replaces FluxKontext dev’s CLIP-based encoders with Qwen-VL-7b embeddings for image and instruction conditioning.The replacement is intended to improve semantic alignment.
- Experimental Setup: The experiments cover general editing with GEdit-EN-full and ImgEdit-Full, compositional understanding with Complex-Edit, and context-aware editing with OmniContext.
- Main Results: 7.236 on GEdit-EN-full, 3.80 on ImgEdit-Full, and 8.78 on Complex-Edit are reported for the model trained on GPT-IMAGE-EDIT-1.5M.On Complex-Edit C8, the score includes 8.99 Instruction Following and 8.41 Identity Preservation.
- Ablation Studies: Output-image regeneration improves Flux 1.0 dev’s OmniEdit GEdit-EN score from 4.93 to 5.98, while instruction regeneration raises imgedit from 3.24 to 3.40.The two steps respectively regenerate outputs and rewrite instructions to match the new outputs.
- Ablation Studies: Raw Complex-Edit instructions produce 5.39 on GEdit-EN and substantially impair identity preservation by altering uninstructed image regions.The result indicates that instruction complexity alone is insufficient without high-quality, aligned image pairs.
- Ablation Studies: Including Complex-Edit data improves average scores from 7.03 to 7.24 on GEdit-EN and from 3.71 to 3.80 on ImgEdit.The ablation describes this as a benefit of training on more challenging compositional instructions.
- Qualitative Results: Qualitative results show instruction understanding, realistic image generation, and preservation of elements not specified for change.
5 Conclusion
The work introduces GPT-IMAGE-EDIT-1.5M, a 1.5-million-sample corpus refined around instruction following, identity preservation, and perceptual quality. Fine-tuning on it yields state-of-the-art performance across multiple benchmarks and narrows the gap with proprietary models.
- GPT-IMAGE-EDIT-1.5M contains over 1.5 million samples created by refining and unifying existing image-editing datasets.The curation process emphasizes instruction following, identity preservation, and perceptual quality.
- The qualitative results are presented for Img-Edit, OmniContext, and Complex-Edit.
- Fine-tuning a state-of-the-art open-source model on GPT-IMAGE-EDIT-1.5M achieves new state-of-the-art performance across multiple benchmarks.The conclusion reports substantial gains and a narrower gap with proprietary models.
- Ablation studies show tangible benefits from each refinement step, including output regeneration and instruction rewriting.The reported findings identify tight instruction-image alignment and identity preservation as more important than instruction complexity.
- The paper releases GPT-IMAGE-EDIT-1.5M and fine-tuned models as resources for further open-source image-editing research.Future directions include applying the curation methodology to video or 3D and automating misalignment detection.
A.1 UltraEdit Downscale Workflow
The UltraEdit workflow regenerates images at higher resolution before restoring the benchmark-compatible dimensions. This preserves compatibility while improving the processing resolution.
- UltraEdit inputs were originally 512×512, while the workflow regenerated applicable inputs and outputs at 1024×1024.
- Both regenerated images were downsampled to 512×512 using bicubic interpolation.
- The downsampling step maintained compatibility with the original benchmark’s expectations.
A.2 OmniEdit Alignment Procedure
The OmniEdit alignment procedure normalizes image geometry before editing and restores original dimensions afterward. A strict border-padding filter removes samples with excessive residual padding.
- Each OmniEdit image is padded to the nearest supported generation ratio before editing.The procedure first computes the image ratio and applies geometric alignment.
- After generation, padding is cropped and the image is resized back to its original dimensions.This is intended to ensure comparable pixel density.
- Samples are rejected when more than 0.5% of the border remains uniform padding after processing.
A.3 Complex-Edit Subset
The Complex-Edit subset uses the OmniEdit geometry and padding procedure while retaining only canonical complex instructions. Outputs with detectable padding errors are discarded.
- Complex-Edit follows OmniEdit’s geometry and padding procedure.
- The subset contains only canonical complex instructions.
- Outputs with detectable padding errors after cropping are discarded.
A.4 HQ-Edit Dual Splits
HQ-Edit is processed through separate Edit and Generate splits. Existing pairs retain their original resolution after padded editing, while generation tasks synthesize reference inputs and vary aspect ratios for diversity.
- HQ-Edit is divided into two distinct splits: Edit Split for existing image pairs and Generate Split for generation tasks.
- Edit Split: In the Edit Split, the original image is padded before editing, then cropped and restored to its original resolution.
- Generate Split: In the Generate Split, a new reference input image is synthesized from the textual instruction before applying the same edit instruction.
- Generate Split: Aspect ratios are randomly selected from 1:1, 2:3, and 3:2 to increase diversity in generation tasks.