GPT-IMAGE 1: 4,000 Aesthetically Curated Images with Qwen3-VL Captions

4,000 images were selected for visual clarity, composition, lighting, and stylistic appeal. Each image is captioned using the Qwen3-VL model and provided in PNG format at a resolution of 1024x1536. The dataset was created by user a3xrfgb and last updated on Hugging Face in December 2025.

Use Cases

• Fine-tune text-to-image models based on the high-quality, aesthetically curated images.
• Train or evaluate vision-language models based on the Qwen3-VL generated captions.
• Benchmark image generation quality using a set of images selected for visual clarity and composition.
• Develop aesthetic scoring models based on the described selection criteria for lighting and style.

Strengths

• Contains 4,000 images, providing a substantial collection for model training.
• All images are described as high-quality and curated for aesthetic factors like composition and lighting.
• Each image is paired with a caption generated by the Qwen3-VL model, adding a text modality.
• Images are provided at a consistent, high resolution of 1024x1536 pixels.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The source and original creation method for the 'GPT-IMAGE 1' images are not specified.

Provenance

Source

huggingface

Collection Method

Curated from 'GPT-IMAGE 1' images and captioned with Qwen3-VL.

Freshness

Last updated 2025-12-27 21:51:48; freshness should be verified.