--- title: Qwen-Image-Self-Generated-Dataset canonical_url: "https://www.modelscope.cn/datasets/DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset" md_url: "https://www.modelscope.cn/datasets/DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset.md" repository: DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset last_updated: 2025-08-22 license: "Apache License 2.0" storage_size: "491 GB" domain: - multi_modal tasks: - text-to-image-synthesis downloads: 25006 stars: 27 --- # Qwen-Image-Self-Generated-Dataset > Qwen-Image-Self-Generated-Dataset - DiffSynth-Studio 在 ModelScope 开源的数据集。Generated Image Dataset from Qwen-Image DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset 是 ModelScope 魔搭社区上的text-to-image-synthesis数据集，涉及 multi_modal 领域，存储大小 491 GB，采用 Apache License 2.0 许可。 - **Repository**: DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset - **License**: Apache License 2.0 - **Tasks**: text-to-image-synthesis - **Domain**: multi_modal - **Storage size**: 491 GB - **Downloads**: 25006 - **Stars**: 27 - **Last updated**: 2025-08-22 Source: https://www.modelscope.cn/datasets/DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset --- # Generated Image Dataset from Qwen-Image ## Intruduction This is an image dataset generated using the Qwen-Image model, with a total of 160,000 `1024 x 1024` images, including the general data subset (diffusion_db subset), the English text rendering data subset (en_text_anywordlaion subset), and the Chinese text rendering data subset (zhtext_anywordlaion subset). The input prompts come from DiffusionDB and AnyWord. |||| |-|-|-| |![](https://www.modelscope.cn/datasets/DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset/resolve/master/assets/diffusiondb.png)|![](https://www.modelscope.cn/datasets/DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset/resolve/master/assets/en_text.png)|![](https://www.modelscope.cn/datasets/DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset/resolve/master/assets/zh_text.png)| We provide caption, entity and control images annotations for each image. The annotation model we used is `Qwen2.5-VL-72B-Instruct`. ### Caption Annotations We provide caption annotations for each image in `image_captions.jsonl`. Example: ``` image_path: diffusion_db/000469.png chinese_short_description: 戴着墨镜的皮卡丘。 chinese_long_description: 这是一张展示了一只可爱的皮卡丘的图片，皮卡丘是黄色的，有着标志性的红色脸颊和闪电形状的尾巴。它戴着一副黑色的墨镜，显得非常酷炫和时尚。皮卡丘的表情平静，微微微笑，整体氛围轻松愉快。背景是深色的，突出了皮卡丘的形象，使其更加引人注目。 english_short_description: Pikachu wearing sunglasses. english_long_description: This image features a charming Pikachu, the iconic yellow Pokémon with distinctive red cheeks and a lightning bolt-shaped tail. Pikachu is adorned with a pair of black sunglasses, giving it a cool and stylish appearance. Its expression is calm and slightly smiling, creating a relaxed and cheerful atmosphere. The background is dark, which effectively highlights Pikachu's figure, making it stand out prominently. ``` ### Entity Annotations We provide entity-level annotations for each image in `image_entities.jsonl`. Example: ``` image_path: diffusion_db/000038.png caption: A young woman in a witch hat and orange dress stands with hands clasped, surrounded by vibrant autumn flowers at the hem of her dress. entities: [{"entity": "Young woman in witch attire", "bboxes": [[0.36486486486486486, 0.0472972972972973, 0.6032818532818532, 0.8204633204633205]]}, {"entity": "Vibrant autumn flowers", "bboxes": [[0.24034749034749034, 0.48359073359073357, 0.6959459459459459, 0.9536679536679536]]}] ``` ### Control Image Annotations We provide control images for each image in `control_images.jsonl`. Example: ``` image_path: diffusion_db/053769.png control_images: canny: control_images_new/canny/diffusion_db/053769.png depth: control_images_new/depth/diffusion_db/053769.png softedge: control_images_new/softedge/diffusion_db/053769.png lineart: control_images_new/lineart/diffusion_db/053769.png lineart_anime: control_images_new/lineart_anime/diffusion_db/053769.png openpose: control_images_new/openpose/diffusion_db/053769.png normal: control_images_new/normal/diffusion_db/053769.png ``` ### Text Quality Assessment For the Chinese and English text-render image subsets, we used VLM to perform text quality assessment. Images containing **text that is too small, incorrect text, or overlapping text elements** are considered low-quality. The quality assessment results can be found in the `low_quality_text` field of `image_text_quality.jsonl`. Example: ``` image_path: en_text_anywordlaion/0036144.png low_quality_text: false ``` ## Download ``` pip install modelscope modelscope download DiffSynth-Studio/Qwen-Image-Self-Generated-Dataset --repo-type dataset --local_dir qwen_image_data ``` ## How To Use For annotations, see above. Image files are stored in the files, we provide a script to unzip the images from tar files to subfolders in `./images`: ```python import os import subprocess from tqdm import tqdm import multiprocessing import glob source_tar_dir = "tars" target_image_dir = "images" os.makedirs(target_image_dir, exist_ok=True) def unpack_tar(tar_path, dest_dir): cmd = [ "tar", "xf", tar_path, "-C", dest_dir ] subprocess.run(cmd, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL) return os.path.join(dest_dir, os.path.basename(tar_path).replace('.tar', '')) def unpack_split_tar(split_prefix, dest_dir): cmd = f"cat {split_prefix}* | tar xf - -C {dest_dir}" subprocess.run(cmd, shell=True, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL) return os.path.join(dest_dir, os.path.basename(split_prefix)) def process_tar(args): tar_path, dest_dir = args if '.tar.part_' in tar_path: base_name = os.path.basename(tar_path).split('.tar.part_')[0] split_prefix = os.path.join(os.path.dirname(tar_path), f"{base_name}.tar.part_") return unpack_split_tar(split_prefix, dest_dir) else: return unpack_tar(tar_path, dest_dir) all_tar_files = [] files = os.listdir(source_tar_dir) processed_files = set() for f in files: if f.endswith('.tar.part_aa'): base_name = f.split('.tar.part_')[0] split_prefix = os.path.join(source_tar_dir, f"{base_name}.tar.part_") all_tar_files.append((split_prefix, target_image_dir)) for part_file in glob.glob(f"{split_prefix}*"): processed_files.add(os.path.basename(part_file)) elif f.endswith('.tar') and f not in processed_files: if '.tar.part_' not in f: all_tar_files.append((os.path.join(source_tar_dir, f), target_image_dir)) num_processes = min(50, multiprocessing.cpu_count()) with multiprocessing.Pool() as pool: results = [] print(f"Total TAR files to unpack: {len(all_tar_files)}") with tqdm(total=len(all_tar_files), desc="Unpacking TAR files") as pbar: for res in pool.imap_unordered(process_tar, all_tar_files): pbar.update(1) results.append(res) print(f"\nAll TAR files unpacked to: {target_image_dir}") print(f"Total unpacked folders: {len(results)}") ``` Furthermore, you can verify file correctness using the following code and inject your absolute path into the annotations: ```python import os import random import json from tqdm import tqdm def read_jsonl(file_path, num_samples=None): print(f"reading from {file_path}") data_list = [] samples = 0 with open(file_path, 'r', encoding='utf-8') as file: for line in tqdm(file): data = json.loads(line.strip()) data_list.append(data) samples += 1 if num_samples is not None and samples >= num_samples: break print(f"read {len(data_list)} samples") return data_list root_path = 'path_to_images' datas = read_jsonl('image_captions.jsonl') random.shuffle(datas) for data in datas[:1000]: image_path = data['image_path'] full_path = os.path.join(root_path, image_path) assert os.path.exists(full_path), f"Image does not exist: {full_path}" print(f"All images exists") ```