GPIC: A Giant Permissive Image Corpus (~28 Trillion Pixels) for Visual Generation

Forum topic · 小凯 · 2026-06-01

Summary

GPIC is a newly introduced giant permissive image corpus for visual generation research, containing approximately 28 trillion pixels of diverse internet images. All images are licensed for both research and commercial use, making the corpus uniquely accessible compared to restricted datasets. GPIC is annotated with state-of-the-art vision-language models and is split into 100 million training samples, 200,000 validation samples, and 1 million test samples. The dataset has undergone safety filtering and deduplication, and is centrally hosted on Hugging Face. Alongside the corpus, the authors provide a generation modeling benchmark protocol and a reference baseline using pixel-space flow matching. The team behind GPIC includes researchers from Stanford and collaborators such as Li Fei-Fei, Jiajun Wu, Justin Johnson, Michael Poli, and others. Dataset, benchmark, and models are publicly available on Hugging Face. Paper ID: arXiv 2605.30341.

Paper Overview

Research Area: Computer Vision Authors: Keshigeyan Chandrasegaran, Kyle Sargent, Suchir Agarwal, Michael Jang, Michael Poli, Juan Carlos Niebles, Justin Johnson, Jiajun Wu, Li Fei-Fei Published: 2026-05-28 arXiv: 2605.30341

Abstract (English)

Scalable methods for visual generation modeling require large, accessible, and stable datasets. The authors introduce GPIC, a giant permissive image corpus of approximately 28 trillion pixels. GPIC contains diverse internet images annotated by a state-of-the-art vision-language model, and includes 100 million training samples, 200,000 validation samples, and 1 million test samples. All GPIC images are licensed for both research and commercial use. The corpus has been safety-filtered, deduplicated, and is centrally hosted on Hugging Face. The authors also provide a GPIC generation modeling benchmark protocol and a reference baseline using pixel-space flow matching. The dataset, benchmark, and models are available on Hugging Face.

Key Highlights

Scale: ~28 trillion pixels of diverse internet imagery
Permissive licensing: all images cleared for research and commercial use
Splits: 100M train / 200K validation / 1M test samples
Annotations: generated by a state-of-the-art vision-language model
Curation: safety-filtered and deduplicated
Hosting: centrally hosted on Hugging Face
Benchmark: includes a generation modeling protocol with a pixel-space flow matching reference baseline

---

*Auto-collected on 2026-06-01*