Literature
Display
← All papers

GPIC: A Giant Permissive Image Corpus for Visual Generation

Keshigeyan Chandrasegaran · Kyle Sargent · Suchir Agarwal · Michael Jang · Michael Poli · Juan Carlos Niebles · Justin Johnson · Jiajun Wu · Li Fei-Fei

Preprint posted · arXiv (preprint) ↗

Computer vision7 min text · CondensedOriginal paper ↗
Contents

Overview

The opening gallery retains example GPIC images and their generated captions.

Refer to caption
Figure 1: Example image-caption pairs from GPIC. Additional samples are shown in Figure A.1.

At a glance

GPIC is a centrally hosted image–text corpus with 100 million training, 200,000 validation, and one million test examples, totaling about 28 trillion pixels. The authors collect images from selected permissive source-license categories, filter and deduplicate them, and generate captions with Qwen3-VL-4B. A fixed held-out evaluation protocol and a pixel-space flow-matching baseline accompany the dataset. The paper’s licensing and filtering claims describe its construction process, not an independent audit of every item.

1 Introduction

Modern image generation needs larger and richer text-conditioned datasets than ImageNet-1K, while unstable URL lists and private corpora impede reproducibility. GPIC aims to combine permissive sourcing, fixed hosted files, scale, and straightforward sharded access. It also proposes a less saturated evaluation setup using DINOv2 features and held-out reference statistics. Dataset scale and metric choice are separate contributions; neither alone ensures human-aligned model quality.

Property ImageNet-1K [21] YFCC100M [16] OpenImages [22] DataComp [15] GPIC Permissive ✓\checkmark ✓\checkmark Stable ✓\checkmark ✓\checkmark ✓\checkmark Large ✓\checkmark ✓\checkmark ✓\checkmark Accessible ✓\checkmark ✓\checkmark ✓\checkmark

Table 1: Existing image benchmark datasets fail to satisfy all four criteria. GPIC satisfies all four criteria.

Refer to caption

Figure 2: GPIC dataset statistics. The figure shows GPIC’s image height and width distributions, license composition, caption statistics, release format, dataset splits, and benchmark scales. GPIC images have an average height of 479 pixels and an average width of 587 pixels. GPIC is centrally hosted on Hugging Face as 8,000 shards totaling 12.9TB and released under the MIT license. GPIC-Lite (10M) and GPIC-Nano (1M) provide smaller subsets for development. Best viewed in color.

2 Dataset Construction

Construction proceeds through licensed-source collection, filtering, duplicate removal, captioning, and fixed split packaging.

Original figure
Figure 3: Our dataset construction pipeline. We develop a four-stage pipeline to create GPIC. We source permissive images from Flickr and Wikimedia (Stage 1), filter low-quality and harmful images (Stage 2), deduplicate images using similarity scores derived from SSCD [23] copy detection features (Stage 3), and caption into one of tag, short, medium, or long (Stage 4). Qwen-3-VL-4B-Instruct [19] is used for filtering and captioning.

2.1 Source Pool and Licensing

The initial pool contains 110.57 million Flickr and Wikimedia images, approximately 87.7% and 12.3% respectively. Source categories are CC BY, CC0, Public Domain, and No-Known-Restrictions. Metadata retains license names, license links, attribution, dimensions, and retrieval times, while the release omits retrieved image URLs. Per-image source terms remain relevant alongside the corpus’s stated MIT release license; those are not interchangeable labels.

2.2 Image Filtering

Resolution and aspect-ratio filters remove unsuitable extremes and images whose longest side is under 256 pixels. Qwen3-VL-4B then filters poor-quality or semantically weak images and images it flags as unsafe. The reported quality and safety removals are about 0.3% and 0.35% of the source pool. Automated filtering reduces selected problems but does not establish that every remaining image is free of harmful or sensitive content.

Refer to caption

Figure 4: Example images that are filtered due to low resolution and poor visual quality.

2.3 Deduplication

SSCD copy-detection features and approximate nearest-neighbor search identify similar images. Small-subset collision models help estimate removal rates, but the final rule is not simply a 0.95 threshold. A candidate graph connects pairs above 0.90; pairs above 0.9625 lose the lower-resolution image, and connected components of at least five retain only their highest-resolution image. About 101.3 million remain after deduplication. File-byte SHA-256 checking removes exact byte duplicates, while visually similar items can remain. The conservative choice trades some residual duplication for fewer false removals.

Refer to caption

Figure 5: Qualitative examples of similar image pairs across SSCD similarity ranges. Each group shows nearest-neighbor image pairs within the indicated SSCD similarity interval. At lower thresholds, similar pairs often contain visually related but distinct images, including changes in pose, viewpoint, or object identity. At higher thresholds, pairs increasingly correspond to near-duplicates, but visible differences can still remain (highlighted in red). Together with the high cost of obtaining permissively licensed images at scale, these examples motivate conservative duplicate removal rather than aggressive thresholding. Best viewed in color.

Original figure

Figure 6: Image collision models. SSCD-based duplicate removals follow a power-law trend across subset sizes and similarity thresholds. Extrapolating to the 110M-image source pool shows that θ=0.95\theta=0.95 is estimated to remove 9.62×1069.62\times 10^{6} images, leaving approximately 1.01×1081.01\times 10^{8} images.

2.4 Captioning GPIC

Synthetic captions replace noisy or missing source text. Each image receives one assigned caption format rather than all four formats.

Caption formats.

The mixture is 1% tags, 45% short, 45% medium, and 9% long captions. A 1,520-image benchmark uses human-corrected reference captions and an automated judge scoring summary, count, spatial relation, attribute binding, and text transcription. Qwen3-VL-4B is selected for quality/throughput balance against 2B, 8B, and sparse 30B alternatives. Full captioning reportedly costs about 1,500 H100 GPU-hours. Because the judge compares against reference captions rather than exhaustively inspecting every claim in the image, scores do not measure all possible hallucinations.

Original figure
Figure 7: Captioning model selection. We evaluate Qwen3-VL-Instruct models on the GPIC captioning microbenchmark across five caption-quality criteria and throughput. Throughput in images per second (1xH100) is shown in parentheses below each model. Qwen3-VL-4B-Instruct provides the best quality-throughput tradeoff: it matches or approaches the best quality scores across short, medium, and long captions while maintaining higher throughput than larger models.

2.5 Split Construction and Release

Release design fixes split composition and offers nested scales for development and full training.

Dataset splits.

The released splits contain 100 million training, 200,000 validation, and one million test images. Each preserves the overall source and caption-format proportions. This controls those broad composition axes, rather than guaranteeing balance across semantic or demographic categories.

Benchmark scales.

Nano, Lite, and Full use one, ten, and 100 million training examples. They correspond to the first 80, 800, and 8,000 shards and retain the same source and caption mixtures, enabling nested scaling studies.

Packaging and release.

Images, captions, and metadata are distributed as tar shards, approximately 12,500 images per shard. The full release is reported as 12.9 TB. Frozen files avoid silent changes from recrawling unavailable source links, while approximately balanced sharding simplifies streaming training.

Original figure

Figure 8: GPIC shard statistics. We show the per-shard distribution of image counts and caption-type percentages for GPIC-Full. GPIC-Full is shuffled into 8000 approximately balanced shards, each containing ≈\approx 12,500 images and preserving the target caption mixture of 1% tag, 45% short, 45% medium, and 9% long captions.

3 Benchmarking Protocol

The benchmark compares generated and held-out real image distributions, with a fixed caption set and specified feature extractor.

Metrics.

The primary metric is Fréchet distance over DINOv2 features, supplemented by precision, recall, density, coverage, and maximum mean discrepancy. The paper argues that Inception-based ImageNet FID is saturated for recent models. DINOv2 distance provides another measurement space; it is still a proxy whose scale and rankings depend on the selected representation.

Evaluation protocol.

Models generate 50,000 images from the released fixed captions. Statistics are compared with a one-million-image test reference. The evaluation toolkit and precomputed statistics define the intended reproducible setup; changing prompts or feature extraction changes the comparison.

Held-out test statistics.

Using held-out rather than training-set reference statistics reduces a direct incentive to reproduce the training distribution’s exact samples. This is better separation for evaluation, but it does not by itself detect all memorization or source overlap in pretrained auxiliary models.

Oracle References.

Real-subset comparisons provide context: Test-50K versus Test-1M has distance 7.44, while larger real subsets give lower finite-sample distances. These are empirical references, not a universal zero target. The protocol treats training directly against DINOv2-aligned features or losses as nonstandard and asks for disclosure of auxiliary models, caption rewriting, and other deviations. Matching the evaluation representation during training can improve the score without a corresponding general improvement in generation.

Original figure
Figure 9: Comparison of FID and FD-DINOv2 on ImageNet-1K. ImageNet-1K FID is saturated: several models achieve lower FID than the distance between 50K held-out real ImageNet-1K images and the ImageNet-1K training set. By contrast, FD-DINOv2 remains unsaturated: all evaluated models have higher FD-DINOv2 than the corresponding held-out real-image distance, including models trained with DINOv2 features. Dotted lines indicate the distance between 50K held-out real images and the ImageNet-1K training set. SiD2 [12] is omitted from the FD-DINOv2 comparison because checkpoints or generated samples are not available.
GPIC Subset FD ↓\downarrow Precision ↑\uparrow Recall ↑\uparrow Density ↑\uparrow Coverage ↑\uparrow
Full 1.19 0.947 0.950 1.000 0.972
Lite 1.25 0.951 0.947 1.010 0.973
Nano 1.60 0.946 0.946 1.002 0.968
Val 2.37 0.948 0.949 0.993 0.966
Test-50K 7.44 0.949 0.953 0.997 0.967
Table 2: Oracle reference metrics over DINOv2 features for GPIC subsets evaluated against 1M GPIC Test set. These real-vs-real values provide reference points for interpreting generative model performance on GPIC. Metrics over Inception-v3 representations are provided in Table B.1. Density is not upper bounded by 1.0, so values above 1.0 are valid.

4 Experiments

The reference model is a 1.1-billion-parameter JiT-T2I pixel-space flow matcher conditioned by Qwen3-1.7B. It trains for one epoch at 256×256 with batch size 256, AdamW, and about 40 hours on eight H100 GPUs; streaming errors repeat a small number of samples. Evaluation uses 50 Euler steps. Guidance scales 1.75, 4.0, and 6.25 yield distances 204.01, 87.80, and 76.25; the highest tested scale also improves recall and coverage. This is a reproducible starting point rather than an optimized state-of-the-art model.

Refer to caption

Figure 10: Pretraining loss for the JiT-T2I reference baseline [34] on GPIC-Full. We show training loss vs. iterations. The model is trained for one epoch on GPIC-Full (100M text-image pairs).
Refer to caption
Figure 11: JiT-T2I samples after training on GPIC-Full for one epoch. We show generated images for prompts in the held-out Test-50K subset. Each group contains a real test image, the corresponding text prompt, and JiT-T2I generations sampled with classifier-free guidance scales CFG=1.75,4.00,\mathrm{CFG}=1.75,4.00, and 6.256.25, respectively. The examples span diverse object-centric and scene-level prompts, including animals, vehicles, natural scenes, architecture, and indoor environments. Quantitative results for each CFG scale are reported in Table 3.

CFG FD ↓\downarrow Precision ↑\uparrow Recall ↑\uparrow Density ↑\uparrow Coverage ↑\uparrow 1.75 204.01 0.917 0.530 1.034 0.806 4.00 87.80 0.933 0.765 1.012 0.906 6.25 76.25 0.942 0.792 1.014 0.908

Table 3: JiT-T2I baseline results on GPIC-Full after training for one epoch. We report FD, Precision, Recall, Density, and Coverage for three classifier-free guidance scales. We use 50-step Euler sampling for all generations. All metrics are computed against the 1M GPIC test set.

5 Conclusion

GPIC offers a large hosted corpus and explicit evaluation protocol for open visual-generation research. Remaining limitations include source-platform biases, possible harmful use, memorization, automated-caption errors, residual near-duplicates, and imperfect filtering. Source license metadata improves traceability but does not establish that all potential downstream concerns are resolved. The fixed release and synthetic captions address specific reproducibility and source-text issues, not every privacy or content risk.

Acknowledgments

The authors acknowledge compute support from Radical Numerics and World Labs, and feedback from collaborators and Stanford research groups.

Appendix

The appendices retain additional image–caption examples, evaluation details, filtering and deduplication galleries, caption-correction examples, and all five exact prompt figures.

Refer to caption
Figure A.1: Additional example image-caption pairs from GPIC.

Appendix B Evaluation

Supplementary evaluation specifies preprocessing, real-data reference metrics, and the effect of feature-extractor variants.

B.1 Construction of Imagenet-256 and GPIC-256

Reference images are center-cropped to a square and bicubically resized to 256×256 using Pillow. Interpolation implementations differ across libraries, so matching this detail matters for comparable feature statistics.

B.2 Additional Oracle Reference Metrics

Tables B.1 and B.2 retain Inception-based reference metrics and the mean/covariance contributions to Fréchet distance. The covariance term dominates these real-subset comparisons. Density can legitimately exceed one, so it should not be interpreted as a bounded percentage.

GPIC Subset FD ↓\downarrow Precision ↑\uparrow Recall ↑\uparrow Density ↑\uparrow Coverage ↑\uparrow
Full 0.07 0.757 0.762 0.973 0.966
Lite 0.07 0.762 0.768 1.041 0.973
Nano 0.11 0.756 0.769 1.000 0.973
Val 0.21 0.760 0.758 1.007 0.971
Test-50K 0.68 0.7688 0.763 0.979 0.964
Table B.1: Generative quality metrics over Inception-v3 representations across GPIC subsets against GPIC-Test-1M. We omit MMD as each subset scores ≈0\approx 0.
DINOv2 Inception-v3
GPIC Subset FDμ\mathrm{FD}_{\mu} ↓\downarrow FDΣ\mathrm{FD}_{\Sigma} ↓\downarrow FDμ\mathrm{FD}_{\mu} ↓\downarrow FDΣ\mathrm{FD}_{\Sigma} ↓\downarrow
Full 0.051 1.140 0.005 0.065
Lite 0.052 1.197 0.005 0.069
Nano 0.053 1.547 0.005 0.100
Val 0.012 2.353 0.001 0.208
Test-50K 0.040 7.404 0.004 0.678
Table B.2: FDμ\mathrm{FD}_{\mu} and FDΣ\mathrm{FD}_{\Sigma} across GPIC subsets against GPIC-Test-1M.

B.3 Effect of DINOv2 Backbone Size and Register Tokens on FD

DINOv2 backbone size and register tokens change feature magnitude and therefore the numeric distance scale. Registers lower distances for small, base, and large variants, while giant variants are closer. Across the tested configurations, reported mean Pearson correlation is 0.847 and Kendall concordance 0.795, supporting substantial but imperfect ranking agreement. The default remains non-register ViT-L/14; a dedicated human-alignment comparison is future work.

Original figure
Figure B.1: FD-DINOv2 across DINOv2 model sizes and variants with and without registers.
Original figure
Figure B.2: Distribution of DINOv2 feature values on ImageNet-1K train images.

Appendix C Image Filtering

The additional gallery preserves examples removed for low resolution or poor visual quality. Examples illustrate the filter criteria rather than estimating its error rate.

Refer to caption
Figure C.1: Additional low resolution and poor visual quality image examples.

Appendix D Deduplication

The deduplication gallery preserves image pairs and clusters across similarity tiers. The operational rule retains the highest-resolution representative of qualifying duplicates or large clusters; visually related but distinct images are intentionally kept in ambiguous cases.

Refer to caption
Figure D.1: Qualitative examples of deduplication over similarity score tiers and cluster sizes. All clusters with exact similarity score ≥0.9625\geq 0.9625 are removed, and only clusters of size ≥5\geq 5 are removed for similarity scores in [0.9,0.9625)[0.9,0.9625).

Appendix E Microbenchmark

Humans correct initial captions for 1,520 sampled images, especially counting, spatial relations, and fine details. Those revised captions become references for the model judge. The comparison figure preserves original and corrected text, showing why caption-generator errors remain relevant even with a strong pretrained model.

Refer to caption

Tag:

Model: tennis court, athletic shoes, grass, shadow, player.

Human: tennis court, athletic shoes, grass, shadow, lower body.

Short:

Model: A group of people looks at papers on a wall in an office.

Human: A group of people pins papers and sticky notes to an office wall.

Medium:

Model: A baseball batter in a gray uniform stands at home plate after swinging, with the ball traveling across the field to the right. A catcher in red gear and an umpire crouch behind him, while another player stands near the first-base line and fans fill the stadium seats in the background.

Human: A baseball batter in a gray uniform stands at home plate after swinging, with the ball traveling across the field to the right. A catcher in red gear and an umpire crouch behind him, while another player stands near the third-base line and fans fill the stadium seats in the background.

Long:

Model: Five female volleyball players are positioned on a red and teal court, engaged in a match. Three players in red and blue uniforms are on the left side of the net, with two raising their arms to block a yellow volleyball that is mid-air near the net. One player in a yellow jersey with the number 16 is in a low crouch, facing the net, and another player in a yellow jersey is standing to the right, near the sideline with arms extended. A fifth player in a black and yellow uniform is positioned further right, standing with her feet apart and looking toward the ball. The net is stretched across the center of the court, and a blue vertical post with the words “Olympic Games” visible is on the right side of the net. In the background, spectators are seated on benches along the sidelines, and a blue umpire’s chair is visible on the far left. The court surface is marked with white lines, and the net has a white top band with vertical markings.

Human: Seven female volleyball players are positioned on a red and teal court, engaged in a match. Four players in red and blue uniforms are on the left side of the net, with two raising their arms to block a yellow volleyball that is mid-air near the net. One player in a yellow jersey with the number 16 is in a low crouch, facing the net, and another player in a yellow jersey is standing to the right. A fifth player in a black and yellow uniform is positioned further right, standing with her feet apart and looking toward the ball. The net is stretched across the center of the court, and a blue vertical post with “London 2012” visible is on the right side of the net. In the background, spectators are seated on benches along the sidelines, and a blue umpire’s chair is visible on the far left. The court surface is marked with white lines, and the net has a white top band with vertical markings.

Figure E.1: Full caption comparison between VLM-generated and human-labeled annotations. Red highlights VLM captioning model outputs and blue highlights human annotations. Underlined text indicates the specific differences in counting, spatial relations, and fine-grained visual details.

Appendix F Prompts

All exact caption and judge prompts are retained as original figures. Tags request three to eight short phrases; short captions target roughly 12–25 words; medium captions target two or three sentences and 25–60 words; long captions request five to seven sentences and more visible-text transcription. The judge scores five reference-based axes from zero to two, marking inapplicable axes separately. Several criteria explicitly ignore claims absent from the human reference, which limits their ability to detect extra unsupported details.

Original figure
Figure F.1: Prompt used to generate tag-styled captions for GPIC images.
Original figure
Figure F.2: Prompt used to generate short-length captions for GPIC images.
Original figure
Figure F.3: Prompt used to generate medium-length captions for GPIC images.
Original figure
Figure F.4: Prompt used to generate long-sized captions for GPIC images.
Original figure
Figure F.5: Prompt used to evaluate VLM-generated captions against ground-truth captions in our microbenchmark.
Concept demonstration

Open image ↗