datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
GPT-Image-Edit-1.5M
GPT-Image-Edit-1.5M A Million-Scale, GPT-Generated Image Dataset
๐Arxiv | ๐ Project Page | ๐ปGithub
GPT-Image-Edit-1.5M is a comprehensive image editing dataset that is built upon HQ-Edit, UltraEdit, OmniEdit and Complex-Edit, with all output images regenerated with GPT-Image-1.
๐ฃ News
[2025.08.20] ๐ We provide a script for multi-process downloading. See Multi-process Download.
[2025.07.27] ๐ค We release GPT-Image-Edit, a state-of-the-art image editing model withโฆ See the full description on the dataset page: https://huggingface.co/datasets/UCSC-VLAA/GPT-Image-Edit-1.5M.multi_reference_image_editing
Multi-Reference Instruction-Based Image Editing Dataset
Overview
This dataset contains 20,000 high-resolution image pairs and multi-modal instructions designed for training advanced image-to-image editing models. It combines two complementary example types: 10,000 reference-grounded edits, where structural or stylistic changes are driven by up to three provided visual reference images, and 10,000 occlusion-based inpainting/outpainting edits, where the model mustโฆ See the full description on the dataset page: https://huggingface.co/datasets/molbal/multi_reference_image_editing.ramanv-image-editing
ramanv-image-editing
Image editing dataset for training FLUX.1-Kontext / InstructPix2Pix style models.
Size
592,141 total editing pairs
Sources: ultraedit
Schema
Each shard tar contains {uid}_src.jpg, {uid}_edit.jpg, {uid}_mask.png (where available).
Metadata per record: instruction, prompt, edit_type, caption_before/after, license, sha256.
Licenses
MagicBrush, InstructPix2Pix, Pico-Banana, HumanEdit: CC-BY-4.0
UltraEdit, AnyEditโฆ See the full description on the dataset page: https://huggingface.co/datasets/lingamvamshikrishnareddy/ramanv-image-editing.Text_Guided_Image_Editing
Dataset Card
Dataset in ImagenHub.
Citation
Please kindly cite our paper if you use our code, data, models or results:
@article{ku2023imagenhub,
title={ImagenHub: Standardizing the evaluation of conditional image generation models},
author={Max Ku and Tianle Li and Kai Zhang and Yujie Lu and Xingyu Fu and Wenwen Zhuang and Wenhu Chen},
journal={arXiv preprint arXiv:2310.01596},
year={2023}
}
gpt-image-edit-1-5m-hqeditcharacter_turnaround_sheet_qwen_image_edit_2509_datasetBase images were generated by Qwen Image and I used Wan to do 360 degree rotation. I then took frames from the rotation and concatenated them together using imagemagick.
Mask_Guided_Image_Editing
Dataset Card
Dataset in ImagenHub.
Citation
Please kindly cite our paper if you use our code, data, models or results:
@article{ku2023imagenhub,
title={ImagenHub: Standardizing the evaluation of conditional image generation models},
author={Max Ku and Tianle Li and Kai Zhang and Yujie Lu and Xingyu Fu and Wenwen Zhuang and Wenhu Chen},
journal={arXiv preprint arXiv:2310.01596},
year={2023}
}
qwen-image-edit-magic-brush-results
Qwen-Image-Edit MagicBrush Evaluation Results
This dataset contains evaluation results of Qwen-Image-Edit on the MagicBrush benchmark.
Columns: img_id, turn_index, instruction, source_img, generated_img, target_img, ssim, lpips, clip_score, dino_score
Single-turn Average Metrics (250 samples)
SSIM
LPIPS
CLIP
DINO
0.6443
0.4001
0.8968
0.7415
Multi-turn Average Metrics (248 samples)
SSIM
LPIPSCLIP
DINO
0.5354
0.5080
0.8541
0.6268โฆ See the full description on the dataset page: https://huggingface.co/datasets/manharg25/qwen-image-edit-magic-brush-results.ImageEditingRequestV1ramanv-image-editing-assembledImage-Gen-or-Image-Editing
Image Gen or Image Editing
This dataset is designed for text classification of prompts provided by users. It determines whether a prompt is intended for image generation or image editing.
Subject_Driven_Image_Editing
Dataset Card
Dataset in ImagenHub.
Citation
Please kindly cite our paper if you use our code, data, models or results:
@article{ku2023imagenhub,
title={ImagenHub: Standardizing the evaluation of conditional image generation models},
author={Max Ku and Tianle Li and Kai Zhang and Yujie Lu and Xingyu Fu and Wenwen Zhuang and Wenhu Chen},
journal={arXiv preprint arXiv:2310.01596},
year={2023}
}
Text_Guided_Image_Editing_Base64image-editing-style-instruction-following
Image Editing Style Instruction-Following Dataset
A multimodal dataset for evaluating and training models on style-guided image editing via natural language instruction-following. Each sample contains a reference style image, human-written editing instructions, and multiple style-transferred output images.
Overview
Item
Details
Samples
19 sets
Images per sample
1 input (style reference) + 3 outputs (style-transferred)
Total images
76 (19 inputs + 57โฆ See the full description on the dataset page: https://huggingface.co/datasets/obaydata/image-editing-style-instruction-following.pb-image-editing-10k-sftcharacter_turnaround_sheet_qwen_image_edit_2509_datasetBase images were generated by Qwen Image and I used Wan to do 360 degree rotation. I then took frames from the rotation and concatenated them together using imagemagick.
precise_benchmark_for_object_level_image_editing
VOCEdits: A benchmark for precise geometric object-level editing
Sample format: (input image, edit prompt, input mask, ground-truth output mask, ...)
Please refer to our paper for more details: "๐ POEM: Precise Object-level Editing via MLLM control", SCIA 2025.
How to Evaluate?
Before evaluation, you should first generate your edited images.
Use datasets library to download dataset. You should only use input image, edit prompt, and id columns to generate edited images.โฆ See the full description on the dataset page: https://huggingface.co/datasets/monurcan/precise_benchmark_for_object_level_image_editing.MMH3_Image_Edit_WorkflowThis is just an example of using MiniMax H3 as an image editor. The actual workflow that I use requires several custom nodes, some of which are not published, so this one is simply a bare bones demonstration.
This uses the hybrid MiniMax H3 model from here: https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/tree/main
It uses the custom VAE from here: https://huggingface.co/Mamad8/MiniMax-H3-Image-VAE/tree/main
It uses the LoRA from here:โฆ See the full description on the dataset page: https://huggingface.co/datasets/fizzlepoof/MMH3_Image_Edit_Workflow.Image-Editing-ver1gpt-image-edit-benchmark-results
GPT-Image-Edit โ Benchmark Results
This repository contains evaluation results of GPT-Image-Edit across four standard image-editing benchmarks. All scores were computed using the official evaluation scripts provided by each benchmark.
๐ Benchmarks
Benchmark
Metrics
Folder
GEdit-EN
12 editing categories + Avg
gedit/
Complex-Edit
IF, IP, PQ, Overall
complex_edit/
ImgEdit-Full
10 editing operations + Overall
imgedit/
OmniContext
Contextual edit scoresโฆ See the full description on the dataset page: https://huggingface.co/datasets/UCSC-VLAA/gpt-image-edit-benchmark-results.68750-Sets-of-Image-Editing-Reasoning-Based-Editing-Data
Description
68,750๊ฐ ๊ท๋ชจ์ ์ถ๋ก ๊ธฐ๋ฐ ์ด๋ฏธ์ง ํธ์ง(Reasoning-Based Editing) ๋ฐ์ดํฐ์
์ผ๋ก, ์ง์ ๊ธฐ๋ฐ ํธ์ง 62,500๊ฐ์ ๋ฌผ๋ฆฌ ๊ท์น ๊ธฐ๋ฐ ํธ์ง 6,250๊ฐ๋ก ๊ตฌ์ฑ๋ฉ๋๋ค. ๋ค์ํ ์ฅ๋ฉด, ์นดํ
๊ณ ๋ฆฌ ๋ฐ ๋ฌธ์ ์ ํ์ ํฌํจํ๋ฉฐ, ๊ฐ ๋ฐ์ดํฐ๋ ์๋ณธ ์ด๋ฏธ์ง, ํธ์ง ๊ฒฐ๊ณผ ์ด๋ฏธ์ง, ์ค๊ตญ์ด ํ
์คํธ ๋ฌธ์ ๋ฐ ์์ด ํ
์คํธ ๋ฌธ์๋ก ๊ตฌ์ฑ๋ฉ๋๋ค. ๋๋ถ๋ถ์ ์นดํ
๊ณ ๋ฆฌ๋ ์ ๊ณต๋ ์ง์์ ๋ฐ๋ผ ํธ์ง ๊ณผ์ ์์ ์ถ๋ก ์ด ํ์ํ ๋ฐ์ดํฐ๋ฅผ ํฌํจํฉ๋๋ค.
์๋ณธ ์ด๋ฏธ์ง๋ ๋ฌธ์์ ์ ๊ณต๋ ์ง์์ ๋ฐ๋ผ ํธ์งํ์ฌ ๊ฒฐ๊ณผ ์ด๋ฏธ์ง๋ฅผ ์์ฑํ์ผ๋ฉฐ, ์ด๋ฏธ์ง์ ํ
์คํธ ๋ด์ฉ์ ๋งค์นญ ์ ํ๋๋ 95% ์ด์์
๋๋ค. ๋ณธ ๋ฐ์ดํฐ์
์ ๊ฐ์ ์ฅ๋ฉด ์์ฑ, ์ด๋ฏธ์ง ํฉ์ฑ, ๋ฐ์ดํฐ ์ฆ๊ฐ ๋ฐ ์ถ๋ก ๊ธฐ๋ฐ ์ด๋ฏธ์ง ํธ์ง ๋ฑ์ ์์
์ ํ์ฉํ ์ ์์ต๋๋ค.
์์ธํ ๋ด์ฉ์ ์๋ ๋งํฌ๋ฅผ ์ฐธ๊ณ ํด ์ฃผ์ธ์: https://ko.nexdata.ai/datasets/llm/2125?source=Hf.krโฆ See the full description on the dataset page: https://huggingface.co/datasets/Nexdata-kr/68750-Sets-of-Image-Editing-Reasoning-Based-Editing-Data.qwen_image_layered_editing_with_qwen_image_edit_magic_brushmulti-modal-image-editapply_texture_qwen_image_edit_2509_datasetAll input images were created using Qwen-Image (w/ lightning lora) and the outputs were generated using Qwen-Image-Edit-2509. No external sources were used to generate this dataset.
ramanv-image-editing-pairsramanv-image-real-style-editorialmulti-edit-image-pairs
Image Editing Dataset
This dataset contains image editing examples with instructions.
Dataset Structure
instruction: Text instruction for editing
original_image: Original image before editing
edited_image: Image after applying the edit
1.51-Million-Single-Image-Editing-Sample-Data
1.51-Million-Sets-of-Single-image-and-Multi-image-Fusion-Image-Editing-Data
Description
This dataset is just a sample of 1.51 Million Sets of Single-image and Multi-image Fusion Image Editing Data. Editing types include 500,000 sets of portrait/object consistency editing, 300,000 sets of structural edits, 210,000 sets of mixed editing, and 450,000 sets of spatial editing, and 50,000 sets of style transfer editing. The editing targets cover scenes such as people, animalsโฆ See the full description on the dataset page: https://huggingface.co/datasets/Nexdata-AI/1.51-Million-Single-Image-Editing-Sample-Data.pb-image-editing-10k-sftrestoredit-qwen-image-edit