Fei-Fei Li, the Stanford researcher who created ImageNet in 2009, is now building its successor. Called the Giant Permissive Image Corpus (GPIC), the new dataset contains 100 million image-text pairs and approximately 28 trillion pixels, designed specifically for modern generative AI systems rather than the image classification tasks that dominated when ImageNet launched.

Why Is ImageNet No Longer Enough for Modern AI?

ImageNet revolutionized computer vision by providing a massive labeled dataset of images organized into categories. When researchers at the University of Toronto trained a model called AlexNet on ImageNet using NVIDIA GPUs in 2012, the results were so dramatic that the entire field shifted almost overnight. For more than a decade, nearly every major breakthrough in visual recognition was measured against ImageNet's benchmarks.

But the problem facing AI researchers today is fundamentally different. Modern systems don't just classify images; they generate them from text descriptions, create photorealistic scenes, produce artwork and videos, and learn from massive image-text pairs. Evaluating these generative systems using benchmarks designed for classification is like using a 1990s exam to measure skills that didn't exist back then.

The situation has become so acute that some evaluation metrics are now producing nonsensical results. Several recent papers have reported scores on the Fréchet Inception Distance (FID), a common metric for image quality, that actually outperform the scores achieved by real images themselves. When synthetic images score "better" than real images according to the benchmark, the benchmark has stopped measuring what it's supposed to measure.

What Makes GPIC Different From ImageNet?

GPIC represents a complete rethinking of how AI researchers should build and evaluate visual datasets. The project includes researchers from Stanford University, with Fei-Fei Li serving as a leader within Stanford's AI research ecosystem while also founding the spatial intelligence company World Labs.

The dataset itself is massive in scale and carefully constructed. It contains approximately 12.9 terabytes of data organized into 8,000 streaming-ready shards, with 200,000 validation samples and 1 million test images. But size alone isn't what sets it apart. GPIC was designed around several core principles that address problems plaguing modern AI research:

• Legal Clarity: All images come from sources with clearly permissive licenses, including Creative Commons BY, CC0, public domain, and images with no known restrictions. This provides stronger legal clarity than many existing datasets while enabling both academic and commercial research.
• Quality Filtering: Researchers used advanced vision-language models to identify and remove extremely low-resolution images, severely blurred content, overexposed images, nearly blank images, and unsafe content. Additionally, more than one million duplicate images were removed using copy-detection frameworks to ensure higher dataset diversity.
• Rich Metadata: Rather than relying on generic file names or incomplete descriptions, GPIC includes entirely new captions generated for every image at four levels of detail: tags, short descriptions, medium descriptions, and long descriptions. Generating these captions required approximately 1,500 NVIDIA H100 GPU hours.

How Does GPIC's New Evaluation System Work?

Beyond the dataset itself, GPIC introduces a fundamentally new way to evaluate image generation models. The old FID metric relied on feature representations extracted from Inception-v3, a classification network introduced in 2015. The problem is that Inception-v3 was never designed to evaluate generated images, and researchers have increasingly observed situations where lower FID scores don't correspond to better visual quality and models learn to optimize specifically for the metric rather than improve actual image generation.

GPIC replaces Inception-based features with representations derived from DINOv2, Meta's self-supervised vision model. This new metric, called FD-DINOv2, offers stronger semantic understanding, better feature representations, improved alignment with human perception, and greater robustness for visual similarity evaluation.

The evaluation process itself is also more rigorous. Many existing benchmarks compare generated images against training distributions, which means a model can achieve impressive scores simply by memorizing training examples rather than learning meaningful generalizations. GPIC addresses this by evaluating against an independent one-million-image test set, which reduces the risk of benchmark overfitting and enables fairer comparisons across different research teams.

Why Does This Matter for AI Research?

The stakes are surprisingly high. At Stanford, GPU clusters have evolved from experimental tools to essential infrastructure for nearly every significant AI research question the university is pursuing. Researchers cannot compete on some problems if they lack sufficient computing power, and the choice of what benchmarks to use directly influences what research gets funded and pursued.

Fei-Fei Li's involvement in creating GPIC is particularly significant because she helped establish the benchmark infrastructure that defined computer vision research for more than a decade with ImageNet. In many ways, GPIC can be viewed as a successor project from the same research lineage that originally brought ImageNet to the world. By creating a new standard designed for generative AI rather than classification, she and her collaborators are essentially resetting the field's measuring stick at a moment when the old one has become unreliable.

Initial experiments indicate that current generation models remain well below the theoretical ceiling of FD-DINOv2, suggesting the metric retains substantial room for future progress and won't immediately saturate like older benchmarks. This means GPIC could provide a stable foundation for measuring genuine advances in image generation for years to come.