Fei‑Fei Li’s Team Unveils GPIC: A 100‑Million‑Pair Image‑Text Corpus to Supersede ImageNet

In 2012 AlexNet’s victory on ImageNet launched the deep‑learning era, and for more than a decade ImageNet served as the standard benchmark for computer‑vision models such as VGG, ResNet, and ViT. Recent generative‑model papers report that their FID scores on ImageNet are lower than those of real images, indicating that the benchmark is saturated and no longer reflects true model quality.

To address this, researchers from Stanford—including Fei‑Fei Li, Jia‑Jun Wu, and students—released GPIC (Giant Permissive Image Corpus), a massive open‑access dataset designed for visual generation. GPIC contains 100 million training images, 20 000 validation images, and 1 million test images, totaling about 12.9 TB and 28 trillion pixels, hosted on Hugging Face for free download.

The construction follows four strict stages:

Authorized collection only: Images are sourced exclusively from Flickr and Wikimedia under CC BY, CC0, public‑domain, or other clear‑license categories, yielding 110 million raw images (87.7 % Flickr, 12.3 % Wikimedia).

Low‑quality and harmful‑content filtering: Using the visual‑language model Qwen3‑VL‑4B, the team automatically removes low‑resolution, blurry, over‑exposed, near‑blank, or unsafe images, discarding roughly 0.3 % and 0.35 % of the data respectively.

Deduplication: A duplicate‑detection model called SSCD computes pairwise feature similarity; a conservative policy deletes high‑confidence duplicates, leaving about 1.013 million unique images.

High‑quality caption generation: Qwen3‑VL‑4B generates four levels of textual descriptions (label, short, medium, long) for each image, consuming ~1500 H100‑GPU‑hours for the 100 million captions.

GPIC also introduces a new evaluation protocol. The traditional FID metric relies on Inception‑v3, a classifier not optimized for generative quality, and can be gamed. GPIC adopts FD‑DINOv2, which uses Meta’s 2023 self‑supervised DINOv2 model whose feature space aligns better with human perception. Experiments show that all current major generative models still score higher than real images on FD‑DINOv2, indicating headroom for the metric.

Another key improvement is that GPIC’s benchmark scores are computed against an independent million‑image test set rather than the training set, preventing models from simply memorizing training data to achieve high scores.

For reproducibility, the authors trained a reference baseline on GPIC‑Full (1 billion images) using the JiT (Just image Transformers) architecture with a 1.1 B‑parameter transformer backbone. Training on a single node with eight H100 GPUs for ~40 hours (one epoch) at 256×256 resolution yields an FD‑DINOv2 score of 76.25, which, while modest, provides a common starting point for future work.

GPIC is released in three sizes—GPIC‑Nano (1 M images), GPIC‑Lite (10 M images), and GPIC‑Full (100 M images)—to accommodate varying compute resources. The authors argue that an open, reproducible benchmark is essential for transparent progress in visual generation, mirroring the standardization seen in NLP with GLUE and SuperGLUE.

Paper: "GPIC: A Giant Permissive Image Corpus for Visual Generation" (arXiv:2605.30341). Project site: https://gpic.stanford.edu/.