GPIC: A Giant Permissive Image Corpus for Visual Generation

1 Introduction

• Modern visual generative models use much larger, more diverse corpora and rich text conditioning, making class-conditional ImageNet-1K less representative of current practice.The paper also notes that ImageNet-1K FID has saturated and that some methods score lower than held-out real images.Source: §1 · p. 1
• Proprietary or unstable datasets hinder reproducibility and open scientific comparisons, motivating a benchmark dataset designed for modern visual generation.The paper identifies permissiveness, stability, scale, and accessibility as its four criteria.Source: §1 · p. 1, Table 1 · p. 3
• GPIC combines 27.97 trillion pixels with 100M training, 200K validation, and 1M test examples, plus a new FD-DINOv2 protocol and pixel-space flow-matching baseline.The corpus is centrally hosted on Hugging Face in 8,000 shards.Source: §1 · p. 3
• GPIC’s images average 479 pixels high by 587 pixels wide, and the corpus is released under MIT in 8,000 shards totaling 12.9TB.GPIC-Lite and GPIC-Nano provide 10M- and 1M-image subsets for development.Source: Figure 2 · p. 3

2 Dataset Construction

• The four-stage pipeline sources permissive images, filters low-quality and harmful content, removes duplicates, and generates captions in four formats.Qwen-3-VL-4B-Instruct is used for filtering and captioning.Source: Figure 3 · p. 4

2.1 Source Pool and Licensing

• The source pool contains 110,569,761 images from Flickr and Wikimedia, restricted to licenses permitting redistribution and commercial use.License and attribution metadata are retained, while retrieved image URLs are excluded.Source: §2.1 · p. 4

2.2 Image Filtering

• Filtering removes extreme resolutions and aspect ratios, images with a longest side below 256 pixels, poor-quality or semantically limited images, and images flagged as unsafe.The quality and safety stages remove approximately 0.3% and 0.35% of the source pool, respectively.Source: §2.2 · p. 4

2.3 Deduplication

• Conservative SSCD-based deduplication removes lower-resolution images above 0.9625 similarity and keeps only the highest-resolution image in components of at least five, leaving 101.3M images.The approach is calibrated to avoid removing visually distinct images; exact duplicates are checked with SHA-256 hashes.Source: §2.3 · pp. 4–5

2.4 Captioning GPIC

• Qwen3-VL-4B-Instruct offers the selected quality-throughput tradeoff, with strong summary quality, top spatial and attribute-binding scores, and 56.10 and 49.31 images/sec on short and medium captions.Its summary score is 1.68 versus 1.73 for 30B-A3B; captioning the full corpus required approximately 1,500 H100 GPU-hours.Source: §2.4 · p. 6

2.5 Split Construction and Release

• GPIC is split into 100M training, 200K validation, and 1M test images, with Nano, Lite, and Full tiers of 1M, 10M, and 100M images.The dataset is released as 8,000 balanced shards of approximately 12,500 images, preserving the target caption mixture.Source: §2.5 · pp. 6–7, Figure 8 · p. 7

Metrics

• FD-DINOv2 is the primary GPIC metric; unlike ImageNet-1K FID, it remains unsaturated against held-out real-image distances.The protocol also reports Precision and Density for fidelity, and Recall and Coverage for diversity.Source: §3 · p. 7, Figure 9 · p. 8

Evaluation protocol

• Evaluation compares 50K generated images for fixed test captions against reference statistics from the held-out 1M-image GPIC test set.GPIC releases the reference statistics and an evaluation suite with FD-DINOv2, Precision, Recall, Density, Coverage, and Maximum Mean Discrepancy.Source: §3 · p. 7

Oracle references

• Real-versus-real distances from GPIC subsets to Test-1M provide reference points, with Test-50K versus Test-1M for monitoring saturation.Table 2 reports oracle metrics over DINOv2 features; density values above 1.0 are valid.Source: §3 · pp. 7–8, Table 2 · p. 8

Protocol compliance

• DINOv2 features or objectives aligned to FD-DINOv2 are discouraged because feature overlap and representation-matched training make gains difficult to interpret as generative improvement.Such results are treated as non-standard GPIC results.Source: §3 · p. 8

Reporting

• Report auxiliary networks trained on other datasets and protocol deviations, including prompt rewriting and different captioning or text-embedding models.Transparent reporting supports interpretability and comparison across methods.Source: §3 · p. 8

Baseline

• The reference is JiT-T2I, a 1.1B-parameter pixel-space flow-matching Transformer trained on GPIC-Full to provide a reproducible comparison point.JiT uses single-stage training without tokenizer pretraining or auxiliary losses, with Qwen3-1.7B for text conditioning.Source: §4 · p. 9

Experiment setup

• Training used one epoch at 256 × 256 resolution, and evaluation followed the GPIC protocol with 50-step Euler sampling at CFG scales 1.75, 4.0, and 6.25.Training took approximately 40 hours on one 8×H100 node.Source: §4 · p. 9

Training and samples

• Figure 10 plots pretraining loss over iterations, while Figure 11 pairs held-out Test-50K prompts and real images with generations at three CFG scales.The qualitative examples span object-centric and scene-level prompts.Source: Figure 10 · p. 9, Figure 11 · p. 9

Results

• FD reaches its best reported value, 76.25, at CFG 6.25; increasing CFG also improves recall and coverage in this baseline.The authors release the model as a reference for future GPIC comparisons.Source: §4 · p. 10, Table 3 · p. 10

5 Conclusion

• GPIC is a large-scale, high-quality image-text resource for visual generative modeling and broader multimodal research.The paper describes its design choices, construction pipeline, release format, evaluation protocol, compliance guidelines, and reference baseline.Source: §5 · p. 10
• The dataset retains license and attribution metadata, uses generated captions rather than scraped alt text, and is released as frozen tar shards.These choices are described as steps to mitigate societal risks and avoid source-text exposure, silent dataset drift, and URL-level data poisoning.Source: §5 · p. 10
• GPIC carries risks including harmful-generation misuse, training-content memorization, and amplification of source-platform biases.Despite deduplication efforts, the authors estimate that a small prevalence of near-duplicates may remain.Source: §5 · p. 10

Appendix

• The appendix provides additional GPIC image-caption examples.Source: Figure A.1 · p. 15

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

• The protocol center-crops along the longer edge to form a square image.Source: §B.1 · p. 16
• The protocol downsamples images to 256 × 256 using Pillow's bicubic interpolation.Source: §B.1 · p. 16
• Pillow was chosen because its bicubic interpolation is consistent with prior work.Source: §B.1 · p. 16

B.2 Additional Oracle Reference Metrics

• Table B.1 reports generative-quality metrics over Inception-v3 representations for GPIC subsets against GPIC-Test-1M.Source: Table B.1 · p. 16
• Table B.1 omits MMD because each subset scores approximately zero.Source: Table B.1 · p. 16
• Table B.2 reports FDµ and FDΣ for GPIC subsets against GPIC-Test-1M.Source: Table B.2 · p. 16

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

• The ablation tests variants with and without registers across ViT-S/14, ViT-B/14, ViT-L/14, and ViT-g/14.Registers were introduced to reduce high-norm artifacts in DINOv2 feature maps.Source: §B.3 · p. 16
• For small, base, and large backbones, register variants produce lower FD-DINOv2 scores; at giant scale, the variants are much closer.Source: §B.3 · p. 16, Figure B.1 · p. 17
• For small, base, and large backbones, register variants have narrower feature-value ranges; the ViT-g/14 distributions nearly overlap, matching the smaller FD difference.Because Fréchet distance depends on feature means and covariances, changes in range and variance affect its scale.Source: §B.3 · p. 16, Figure B.2 · p. 17
• 0.847 mean pairwise Pearson correlation and 0.795 Kendall’s coefficient of concordance indicate broad evaluator agreement across eight configurations; non-register ViT-L/14 is the default.The authors leave a dedicated human-alignment study of variants with registers to future work.Source: §B.3 · p. 17

C Image Filtering

• Figure C.1 provides additional examples of low-resolution and poor-visual-quality images.Source: Figure C.1 · p. 18

• All clusters with exact similarity scores ≥0.9625 are removed.Source: Figure D.1 · p. 18
• For similarity scores in [0.9, 0.9625), only clusters of size ≥5 are removed.Source: Figure D.1 · p. 18
• Figure D.1 shows qualitative deduplication examples across similarity-score tiers and cluster sizes.Source: Figure D.1 · p. 18

E Microbenchmark

• 1,520 images sampled from the initial source-pooling stage were human-relabeled to correct caption errors and provide ground truth for evaluating Qwen3-VL-Instruct models.Common errors included counting and spatial relations.Source: E Microbenchmark · p. 19
• Examples contrast model and human captions on tags, actions, baseball spatial relations, and volleyball player counts and visible text.The figure highlights differences in counting, spatial relations, and fine-grained visual details.Source: Figure E.1 · p. 19, E Microbenchmark · p. 19

F Prompts

• The prompts cover tag-style, short-, medium-, and long-form captioning, plus comparison of generated captions against ground truth.Source: F Prompts · p. 20, Figure F.1 · p. 21, Figure F.2 · p. 22
• The judge scores five caption criteria from 0 to 2, comparing semantic content with the human caption rather than writing style or length.It accepts natural-language or unordered tag-style captions, permits extra correct information, and penalizes omitted human-caption content.Source: F Prompts · p. 25
• Counting is evaluated only when the human caption gives quantities, with exact agreement scoring 2; spatial understanding is not applicable without a human-caption spatial description.Added counts absent from the human caption are ignored.Source: F Prompts · p. 25
• Attribute binding and visible-text transcription are scored only for attributes or text mentioned in the human caption, with exact matches scoring 2.Additional VLM attributes are ignored; OCR case, punctuation, and spelling must match.Source: F Prompts · p. 25