The 28-trillion-pixel corpus is fully permissive for commercial use.
Jing Yu Koh#1204
1/ Introducing GPIC: a Giant Permissive Image Corpus and benchmark for visual generation!
🚀100M VLM-captioned image-text pairs for training 📊1M image-text pairs for benchmarking 🖼️~28 trillion pixels 🤗Centrally Hosted ✅Fully permissive for research + commercial use
Dataset, benchmark and models🧵👇
Co-led with @KyleSargentAI
Many users praised the GPIC 100M permissive image-text dataset release as a valuable open contribution that enables better training and benchmarking for visual AI models.
No Digg Deeper questions have been answered for this story yet.
I’m very excited by this new benchmark dataset for visual generation that is suitable for the modern era of large scale generative models!🤩
1/ Introducing GPIC: a Giant Permissive Image Corpus and benchmark for visual generation!
🚀100M VLM-captioned image-text pairs for training 📊1M image-text pairs for benchmarking 🖼️~28 trillion pixels 🤗Centrally Hosted ✅Fully permissive for research + commercial use
Dataset, benchmark and models🧵👇
Co-led with @KyleSargentAI
1/ Introducing GPIC: a Giant Permissive Image Corpus and benchmark for visual generation!
🚀100M VLM-captioned image-text pairs for training 📊1M image-text pairs for benchmarking 🖼️~28 trillion pixels 🤗Centrally Hosted ✅Fully permissive for research + commercial use
Dataset, benchmark and models🧵👇
Co-led with @KyleSargentAI
GPIC should be the new standard benchmark for generative modeling. Training 1 epoch on GPIC is the same cost as 100 epochs on ImageNet, but is a much better proxy for real-world problems. If you work in generative modeling, try GPIC for your next project!
1/ Introducing GPIC: a Giant Permissive Image Corpus and benchmark for visual generation!
🚀100M VLM-captioned image-text pairs for training 📊1M image-text pairs for benchmarking 🖼️~28 trillion pixels 🤗Centrally Hosted ✅Fully permissive for research + commercial use
Dataset, benchmark and models🧵👇
Co-led with @KyleSargentAI
Today we released “GPIC: A Giant Permissive Image Corpus for Visual Generation.” It’s a 100M image dataset for visual generation, with text captions and 100% known+permissive licenses, hosted on HuggingFace. I’m excited to get this out! Check it out: https://gpic.stanford.edu/
7/ Happy pretraining! 🤗 Dataset: https://huggingface.co/datasets/stanford-vision-lab/gpic 🤗 Models: https://huggingface.co/stanford-vision-lab/gpic-baselines 🛠️ Code + evaluation toolkit: https://github.com/keshik6/gpic 🌎 Website: https://gpic.stanford.edu 📄 Paper: https://arxiv.org/abs/2605.30341
In recent years, academic and industry work in generative modeling has drifted so far apart that they are playing totally different games, and techniques that work in academia may not transfer to industry problems.
The divide isn't just about scale -- the different tasks in academia vs industry lead to different fundamental challenges.
Academic work focuses on class-conditional ImageNet generation. This has a very weak conditioning signal (single categorical label) and the problem is very data-constrained, with all SOTA methods training for hundreds of epochs. The main challenge in this regime is combatting overfitting.
Industry work on image or video generation usually has a much richer conditioning signal (e.g. very long captions, input images, etc) and is almost always underfitting since data can be scaled to absurd degrees. Overfitting (at least for pretraining) isn't a concern; instead we want to fit the complex data distribution *as fast as possible*.
We hope that GPIC is approachable on the academic budgets people are already expending on ImageNet, but will lead to problems more similar to the industry-scale challenges in generative modeling.
GPIC should be the new standard benchmark for generative modeling. Training 1 epoch on GPIC is the same cost as 100 epochs on ImageNet, but is a much better proxy for real-world problems. If you work in generative modeling, try GPIC for your next project!
2/📌Why does this matter? ImageNet-1K drove progress in visual generation for over a decade. But to train and benchmark modern visual generative models, we need large, permissive, and accessible datasets with rich text captions.
GPIC is designed for this new setting.
4/📊GPIC Statistics: 🚀100M training image-text pairs 🧪1M test + 200K validation pairs 📦12.9TB across 8,000 shards 🤗Centrally hosted on Hugging Face 📝Captions: 1% tag, 45% short, 45% medium, 9% long ⚡ Benchmark scales: GPIC-Nano (1M), GPIC-Lite (10M), and GPIC-Full (100M)
I’m also proud of this section of the paper, which gives best practices for compliance with our eval protocol. Without calling out anyone in particular, let me just say that using auxiliary foundation models to get a better FD-DINOv2 on GPIC without being very up front about the huge advantages of the extra data and model FLOPs is super bad – please don’t do it!
5/🔧GPIC improves the evaluation protocol for visual generative modeling. ImageNet-1K FID is saturated: several models now score better than held-out real images.
GPIC uses FD-DINOv2, an unsaturated metric, evaluated against the 1M GPIC test set.
6/🎯We also provide a reference baseline. We train a 1.1B JiT-T2I pixel-space flow matching model on GPIC for one epoch.
It obtains FD=76.3 on the GPIC benchmark (cfg=6.25).
3/🤔What are the requirements for a visual generation benchmark dataset?
✅Permissive: research + commercial use ✅Stable: the benchmark should not change over time ✅Large: much larger than ImageNet-1K ✅Accessible: centrally hosted and easy to download
GPIC satisfies all four criteria.
@keshigeyan Nice!
Has been looking forward to such resources for so long! Amazing @keshigeyan
Also really like the quote of Goodhart’s law: When a measure becomes a target, it ceases to be a good measure.
1/ Introducing GPIC: a Giant Permissive Image Corpus and benchmark for visual generation!
🚀100M VLM-captioned image-text pairs for training 📊1M image-text pairs for benchmarking 🖼️~28 trillion pixels 🤗Centrally Hosted ✅Fully permissive for research + commercial use
Dataset, benchmark and models🧵👇
Co-led with @KyleSargentAI
Check out Keshik’s awesome thread for more details! It was great co-leading this project with him. In my thread, I’m going to talk more conversationally about our scientific motivations for this dataset, and what we hope the community will get out of it.
@keshigeyan Wow this is a very valuable contribution! Now that @KyleSargentAI pointed it out it’s really odd we don’t have a standard open T2I training set. Congrats Keshik and Kyle on the release!!
8/ This work was done at @StanfordAILab & @StanfordSVL with the amazing @KyleSargentAI, @agarwal_suchir, @michaeljkjang, @MichaelPoli6, @jcniebles, @jcjohnss, @jiajunwu_cs & @drfeifei 🎉 Big thanks to @huggingface for the support!
One practical example is epoch count – “state-of-the-art” models on ImageNet-1K train for 300-1700 epochs (Fig. credit: PixGen). But that’s not the way you would do things outside of an academic comparison – you’d just go get more data!
@drfeifei This new benchmark dataset could enhance training efficiency for large-scale generative models, potentially reducing compute costs by 30%.
It’s kind of crazy to me that in 2026, we don’t have a huge T2I visual generation benchmark dataset. ImageNet-1K is still the standard, but it’s small and is class-conditional only, so over time the dataset has drifted from the way things are done in practice.
