portex/imagenet-synthetic

Synthetic ImageNet 1K

Augment the ImageNet Dataset with 1,000 realistic, synthetically generated images using the original ImageNet object classes.

Dataset Details

Dataset Description

This dataset includes ~1,000 high-resolution, realistically generated images aligned with ImageNet’s original label classes. Designed for evaluating computer vision model performance post-training or augmenting existing ImageNet datasets, it uniquely addresses dataset limitations by offering synthetic yet highly realistic imagery.

• Curated by: PortexAI
• License: MIT

Dataset Creation & Prompting Strategy

The dataset is created using a careful prompting strategy deliberately crafted to maximize photorealism and contextual relevance.

1. Generate Image Prompt

Each prompt follows a consistent template that:

• Names the object explicitly.
• Places the object in a realistic natural environment.
• Describes the object’s anatomical and functional details.

This structure ensures that each prompt provides enough semantic richness and visual grounding for DALL·E to synthesize plausible images that could mimic real-world photographs.

Generate a prompt that places the following object in a genuine, high-fidelity, photorealistic scenario. Your description should capture not only the natural environment but also the detailed scientific and anatomical essence of the object itself. For example, if the object is a “stingray,” a sample prompt might be:

“A photorealistic image of a stingray that embodies its true biological essence. This flat-bodied batomorph, built on a tough, elastic cartilaginous skeleton, features five distinct ventral gill slits beneath its broad, wing-like pectoral fins. Its mantle-like body, with subtle color gradients and natural textures, is rendered in exquisite detail. The eyes and spiracles are positioned on the top of its head, while its ventrally located mouth and slightly protruding upper jaw capture prey with precision. The stingray glides gracefully over a sandy ocean floor in shallow, crystal-clear turquoise waters, where natural sunlight filters through the water, highlighting both the creature’s intricate anatomy and the serene, authentic underwater environment with realistic coral formations and shifting shadows.”

Your object is: {{object_name}}

2. Generate Image using that Prompt

The generated prompt is passed to the image generation model, Dalle3.

Images are generated using the quality='hd' setting at 1024x1024 sizes.

Use

The images folder includes all image files in .png format. The metadata.csv file includes the corresponding labels for each image.

Dataset Card Contact

Be sure to reach out to [email protected] to access more datasets like this.

We are currently expanding our synthetic data offerings and are working closely with data consumers to match them directly to high-quality real-world, augmented, and synthetic data sources.