acceptodds
Under review as a conference paper at ICLR 2027

GPIC: A Giant Permissive Image Corpus for Visual Generation

Abstract

Studying scalable methods for visual generative modeling requires large, accessible, and stable datasets. We introduce **GPIC**, a Giant Permissive Image Corpus of approximately **28 trillion pixels**. GPIC comprises diverse internet images captioned by a state-of-the-art vision-language model, including 100M training, 200K validation, and 1M test examples. Moreover, all GPIC images are permissively licensed for both research and commercial use. GPIC is safety-filtered, deduplicated, and centrally hosted on Hugging Face. We provide a benchmarking protocol for generative modeling on GPIC. Finally, we provide a reference baseline for pixel-space flow matching on GPIC. Our models, evaluation toolkit, and dataset sample are included in the Supplementary. The full dataset will be publicly released.

Primary area:
datasets and benchmarks
Keywords:
Generative modeling, text-to-image generation, visual generative models, image generation, large-scale datasets, permissive datasets, benchmarking

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bE3oOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces a large-scale text-image dataset under permissive licensing with a specified validation and test set. This dataset can be used to test future works around text-image generation under a controlled setting. Do provide an initial baseline the paper also provides results of a diffusion text-to-image model trained on the dataset.

Critical strengths

The paper observes that there is no generally available large-scale text-image dataset with permissive licensing available. This makes it difficult for generative approaches to be compared fairly, since many approaches train on different kinds of data even if they evaluate on the same test sets (e.g., COCO). The new text-image dataset (scale of 100M pairs) seeks to remedy this by providing a fixed data regime under which different generative approaches can be fairly compared. Since all images come under a permissive license the dataset is hosted directly instead of providing URLs (which is common for other datasets but leads to dataset rot since the URLs might not be available at a later time). The dataset is filtered for NSFW and deduplicated and comes with captions provided, split into long captions, medium captions, and tags. Additionally, a simple diffusion transformer baseline is trained and evaluated on the dataset to provide an initial baseline of the quality.

Critical weaknesses

The dataset is probably not of the highest quality since a) deduplication seems to have been run at a threshold such that not too many images get removed; I think deduplication should be run such that deduplicates are removed, regardless of how many images that removes b) only basic aesthetic quality filtering is run on the dataset c) the captions come from a small VLM

Questions for authors

Besides the dataset, can you release all other metadata you already collected, e.g., different captions, the data you used for deduplication, etc. This metadata might be interesting for people who want to explore additional quality filters and how they affect training outcomes.

Is this dataset supposed to be static or will you update it with better captions etc going forward? A lot of quality improvements in generative models come with better data, so always having the best or a high-quality dataset ready will help make sure that other improvements come from actual methodological approaches and not just improved data processing.

1 comment
AuthorsOpenReview ↗

Response to Reviewer bE3o

We thank the reviewer for their valuable time and constructive feedback. The reviewer (1) raised concerns about deduplication, (2) commented on aesthetic quality filtering, (3) inquired about captioning model selection, (4) requested release of additional metadata, and (5) inquired about future dataset updates. We address each point below.

W1: Deduplication

GPIC removes all exact duplicates, verified using SHA-256 hashes, and high-confidence near-duplicates identified using SSCD copy-detection features (Section 2.3). Through manual inspection, we found that even image pairs with high SSCD similarity (0.75–0.95) can remain visually distinct, differing in pose, viewpoint, or object identity (Figure E.1). Given the limited availability of permissively licensed data on the internet, we adopt conservative deduplication to preserve these distinct images.

 ~

W2: Aesthetic quality filtering

As described in Section 2.2, we remove low-resolution images and those with severe visual defects, including blur, underexposure, and overexposure (Figures 4 and D.1). Our goal is to provide a diverse dataset and benchmark for foundational research in visual generation rather than a dataset optimized for aesthetic preferences. A human audit of 50K randomly sampled images identified 45 images with quality defects (0.09%), indicating a very low residual defect rate.

 ~

W3: Captioning model selection

As explained in Section 2.4, we created a captioning microbenchmark to select a model that balances caption quality and throughput. We compared Qwen3-VL models at 2B, 4B, 8B, and 30B-A3B across five quality criteria (Figure 6, Appendix F). Qwen3-VL-4B achieves comparable overall caption quality to the larger 30B-A3B model (1.68 vs. 1.73), with the highest spatial understanding and attribute binding scores. This quality-throughput tradeoff is important because captioning 100M images requires substantial compute (approximately 1,500 H100 GPU-hours using Qwen3-VL-4B).

To further evaluate caption quality, we measured image–text alignment using Qwen3-VL-Reranker-8B [G], a multimodal reranking model that estimates the probability (from 0 to 1) that an image matches a given text.

Score
GPIC-Nano (1M)0.905
GPIC-Val (200K)0.906
GPIC-Test (1M)0.906

Alignment scores are consistent across all three subsets. We additionally conducted a human audit of 50K randomly sampled image–caption pairs from the final GPIC dataset and identified 620 cases (1.24%) with captioning errors. These errors involve counting and spatial relationships in medium and long captions, while the overall scene descriptions remain accurate.

 ~

Q1: Additional metadata release

We thank the reviewer for the suggestion. We are happy to release additional metadata collected during GPIC construction. As described in Sections 2.1 and 2.5, GPIC already includes per-image licensing, attribution, resolution, and caption metadata. Each GPIC image is assigned one of four caption formats, so one caption is generated per image (Section 2.4). For deduplication, we will additionally release image-pair identifiers and SSCD similarity scores.

 ~

Q2: Static dataset and future updates

GPIC is designed to remain a fixed dataset and benchmark (Sections 1 and 2.5), since changes in training data prevent fair comparisons across generative modeling methods. We agree that improved captions and data filtering may benefit future research; such improvements are best released as separately versioned datasets, leaving the original GPIC benchmark unchanged. Our evaluation protocol recommends reporting any deviations such as changes to captions (Section 3), keeping comparisons interpretable.

We hope the rebuttal addresses all your concerns. We are happy to answer any additional questions.

 ~

References

[G] Li et al. Qwen3-VL-Embedding and Qwen3-VL-Reranker: A unified framework for state-of-the-art multimodal retrieval and ranking. arXiv 2026.

Reviewer_XVkuOpenReview ↗
Rating
3
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GPIC, a large image-text corpus intended to serve as a stable, accessible, and permissively licensed benchmark for visual generative modeling. The dataset contains 100M training examples, 200K validation examples, and 1M test examples, amounting to approximately 28 trillion pixels. The images are collected from Flickr and Wikimedia, filtered for quality and safety, deduplicated using SSCD features, and captioned using Qwen3-VL-4B-Instruct. The dataset is packaged into approximately 8,000 tar shards totaling 12.9 TB, with smaller 1M- and 10M-example subsets provided for development.

The paper also proposes an evaluation protocol centered on FD-DINOv2, together with precision, recall, density, coverage, and MMD. Generated samples are compared against statistics from a held-out one-million-image test set. Finally, the authors train a 1.1B-parameter pixel-space flow-matching baseline for one epoch on the full training set and report a best FD-DINOv2 score.

Critical strengths
  1. Stable and accessible large-scale datasets for text-conditioned image generation are genuinely valuable.
  2. The nested Nano, Lite, and Full configurations could make the benchmark usable across different computational settings.
  3. The paper describes source selection, metadata retention, filtering, SSCD-based deduplication, caption creation, split construction, and release packaging. Exact captioning prompts and parts of the evaluation procedure are included in the appendices.
Critical weaknesses
  1. though the paper describes source selection, metadata retention, etc, how can the user be sure that there is no liability especially in commercial use. there might be metadata changes. there might be conflict from duplicates. The claim is stronger than evidence produced. 100M images are not manually validated, at least statistically significant of them.
  2. Obligation to downstream users is not clearly mentioned for generated artifacts.
  3. As authors pointed out there may be a potential overlap as DINOv2 may have been trained on images with the GPIC test set.
Questions for authors

1.Have the authors conducted a manual audit of image-level licensing and attribution metadata? If so, what sampling design, sample size, error rate, confidence interval, and error categories were observed? 2. How many records have missing, malformed, conflicting, or unverifiable license URLs or attribution strings? 3. How was the 1,520-example captioning microbenchmark sampled? Does it preserve the source, content, resolution, language, and caption-type distributions of the final corpus? 4. How stable are model rankings across GPIC-Nano, GPIC-Lite, and GPIC-Full

1 comment
AuthorsOpenReview ↗

Response to Reviewer XVku

We thank the reviewer for their valuable time and constructive feedback. The reviewer raised questions on (1) licensing, (2) obligations for generated artifacts, (3) DINOv2 overlap, (4) license metadata, (5) captioning benchmark construction, and (6) model rankings across dataset tiers. We address each point below:

W1 + Q1: Licensing and liability for commercial use

GPIC only includes images that Flickr or Wikimedia Commons identifies as CC BY, CC0, Public Domain, or No Known Copyright Restrictions (Section 2.1). For each image, we programmatically retrieve the license name, license URL, and attribution string (including a link to the source page) from its source record, together with the retrieval timestamp. We did not conduct a manual audit, as the retrieved metadata matches each platform's rights designation at collection time by construction.

CC licenses are not revocable once granted, except upon breach by the licensee [H], so later license changes on the source platform do not affect retrieved images. As with any dataset of user-uploaded content, GPIC relies on uploaders' rights designations; cases where an uploader licenses a work they do not own, such as a re-upload, cannot be fully verified at 100M scale. We will provide a removal request form for rights holders to dispute specific images. Users remain responsible for complying with license attribution terms and with rights not covered by CC licenses, such as trademark and privacy rights [H].

[H] Creative Commons. Frequently Asked Questions. https://creativecommons.org/faq/

 ~

W2: Obligations to downstream users for generated artifacts

GPIC is released under the MIT License, and each image retains its original license (Section 2.1). GPIC contains no ShareAlike, NonCommercial, or NoDerivatives images, which are the license conditions that can restrict trained models or their outputs [I]. For model training, we recommend that users cite and link to the GPIC dataset, which contains creator attributions for every image. Creative Commons notes that "for AI model training, attribution could be a simple link to the source of the dataset used to train the model" [I]. We will add these details in the revised version.

[I] Creative Commons. Using CC-licensed works for AI training. https://creativecommons.org/using-cc-licensed-works-for-ai-training-2/

 ~

W3: Potential overlap between DINOv2 training data and the GPIC test set

The potential overlap matters when models are trained with DINOv2 features or FD-DINOv2-related objectives, since such models can directly exploit the evaluation representation. For this reason, the GPIC Evaluation Protocol requires that such use be disclosed and treats it as non-standard (Section 3). All methods are evaluated using the same frozen DINOv2 feature extractor and 1M test-set reference statistics, ensuring a consistent evaluation protocol.

 ~

Q2: Missing or malformed license metadata

Records with missing or malformed license or attribution metadata were excluded during dataset construction. Every record in GPIC therefore includes a license name, license URL, and attribution string with a link to the source page. We will clarify this in the revised version.

 ~

Q3: Sampling of the captioning microbenchmark

The 1,520 microbenchmark images were randomly sampled from the GPIC source pool (Appendix F), so their source, content, and resolution distributions closely match the corpus. The caption-type split of the microbenchmark (47% short, 42% medium, 11% long) follows the corpus distribution (45%, 45%, 9%). We exclude tag captions, which make up only 1% of the corpus and consist of simple keyword lists. All captions are in English, except for transcribed non-English text visible in images.

 ~

Q4: Model rankings across nested dataset tiers.

To test whether results on smaller subsets transfer to GPIC-Full, we train two models from different generative modeling families on GPIC-Nano and GPIC-Full under matched compute per model: JiT (1.1B), our pixel-space flow-matching reference baseline, and LlamaGen-XL (0.78B), a representative autoregressive model trained with the authors' official implementation and tokenizer. LlamaGen-XL is trained for 100 epochs on GPIC-Nano versus 1 epoch on GPIC-Full. For JiT, we use a longer budget of 200 epochs on GPIC-Nano versus 2 epochs on GPIC-Full, as diffusion models benefit from substantially more passes over the data than autoregressive models [J].

ModelNano gFD ↓Full gFD ↓
JiT (1.1B)58.5655.79
LlamaGen-XL (0.78B)166.95147.01

The model ranking is stable across GPIC-Nano and GPIC-Full. Due to rebuttal compute limits, each configuration uses a single run. We will include this experiment in the final version.

We hope the rebuttal addresses all your concerns. We are happy to answer any additional questions.

[J] Prabhudesai et al. Diffusion beats autoregressive in data-constrained settings. NeurIPS 2025.

Reviewer_cZgJOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GPIC, a visual generation dataset containing 100 million training image–text pairs. The authors collect images from Flickr and Wikimedia under specified license categories, apply quality and safety filtering, deduplication, and synthetic captioning, and organize the data into fixed shards and subsets of different sizes. They also provide a DINOv2-based evaluation protocol using a held-out test set and a pixel-space flow-matching baseline to support large-scale, reproducible text-to-image research.

Critical strengths

A valuable resource for community research. The large-scale image–text corpus, license records, fixed splits, and accompanying evaluation tools constitute a useful research resource. Centralized hosting and subsets of different sizes can lower barriers to data access and reproducibility, enabling more consistent comparisons across the community.

Critical weaknesses

The contribution reads more like a technical report on dataset construction. Given the larger datasets and more elaborate construction pipelines already used by leading text-to-image models, the proposed cleaning, captioning, and data organization do not demonstrate clear advantages over established practices. The experiments primarily establish a single-model baseline rather than provide controlled comparisons of data construction choices, leaving the research contribution beyond making the resource available insufficiently established.

Questions for authors
  1. Beyond licensing, stable access, and openness, what distinctive advantages does GPIC offer in content coverage, image–text quality, or training outcomes? Are there dataset comparisons using the same model and training budget that substantiate these advantages?
  2. Which core text-to-image challenges is GPIC intended to address, such as semantic coverage, complex composition, or text rendering, or is it primarily intended as a general-purpose research benchmark? How do the dataset design and experiments support this positioning?
  3. The paper describes overexposure, blur, and safety filtering, but the specific quality criteria and residual defect rates remain unclear. What detection, removal, or correction procedures are used for overexposure/underexposure, black borders, watermarks, and sexual or violent content, and are human audit results available?
  4. Captioning model selection relies mainly on a small-scale evaluation; is there independent human validation of hallucinations, omissions, and image–text mismatches in the final corpus? In particular, when captions describe full images but training uses cropped images, how is consistency with the visible content maintained?
3 comments
AuthorsOpenReview ↗

Response to Reviewer cZgJ (1/2)

We thank the reviewer for their valuable time and constructive feedback. The reviewer raised questions on (1) the purpose and contributions of GPIC, (2) data filtering, and (3) caption quality and crop–caption consistency. We address each point below.

W1 + Q1 + Q2: Purpose, contributions, and target audience of GPIC

GPIC's primary goal is to provide dataset and benchmarking infrastructure for academic research in visual generative modeling. Our contributions are twofold:

(1) A large-scale, reproducible image-text dataset. ImageNet-1K has been widely used for class-conditional generation, but its scale and conditioning differ from modern T2I research. As shown in Table 1, no existing dataset or benchmark satisfies all four criteria: permissive, stable, large, and accessible. URL-based datasets are susceptible to link rot, resulting in different training sets over time and preventing fair comparisons across methods. GPIC provides 100M permissively licensed image-text pairs in fixed, centrally hosted shards, satisfying all four criteria (Table 1, expanded below).

PropertyImageNet-1KYFCC100MOpenImagesDataCompVisual GenomeCC3M / CC12MCOYO-700MShareGPT4VGPIC
Permissive✓✓✓
Stable✓✓✓✓
Large✓✓✓✓✓
Accessible✓✓✓✓✓

(2) A standardized benchmarking protocol. On ImageNet-1K, several models achieve lower FID than the distance between held-out real images and the training set, while FD-DINOv2 remains above the real-image reference for all evaluated models (Figure 8). GPIC uses FD-DINOv2 as its primary metric, provides a fixed set of 50K evaluation captions, and computes reference statistics from a held-out 1M-image test set, unlike ImageNet FID. We also release an evaluation toolkit.

Content coverage. GPIC-Full spans Nature (44.33%), People (40.86%), and Media (14.81%) across a ~1,200-category taxonomy, with closely matched distributions across all splits (Appendix C, Figure C.1).

The three nested training scales (GPIC-Nano, Lite, and Full) support research across different compute budgets. Our reference baseline uses 256×256 resolution, consistent with foundational research in visual generation [A–F]. GPIC images have an average height of 479 pixels and width of 587 pixels, helping keep storage and streaming requirements manageable compared with substantially higher-resolution datasets.

Additional baseline. We chose JiT-T2I as our reference baseline because it is a single-stage, pixel-space model that requires no image tokenizer, providing a simple and reproducible reference point for the community to build on; our code is provided in the supplementary. Nevertheless, to address the reviewer’s concern, we additionally train LlamaGen-XL (0.78B), a representative autoregressive model, on GPIC-Full for one epoch using the authors’ official implementation and tokenizer. This provides reference baselines from two generative modeling families. We will include this baseline in the revised version.

ModelFD ↓Precision ↑Recall ↑Density → 1Coverage ↑
JiT-T2I-XXL (1.1B)76.250.9420.7921.0140.908
LlamaGen-XL (0.78B)147.010.9140.7150.9250.822

GPIC is intended as a general-purpose dataset and benchmark for academic research in visual generation. Its value lies in fixing the training and evaluation setup so that generative modeling methods can be compared fairly, rather than in producing stronger models than other corpora. GPIC is not intended to train models that compete with frontier T2I systems.

AuthorsOpenReview ↗

Response to Reviewer cZgJ (2/2)

Q3: Image quality and safety filtering

We thank the reviewer for the question. We provide the filtering procedures in Section 2.2 and Appendix D.1, including detection criteria, thresholds, removal rules, and qualitative examples (Figures 4 and D.1). The filtering prompts are provided in the supplementary. We summarize our procedures and provide human audit results below.

(1) Image quality filtering. We use Qwen3-VL-4B to identify and remove images with poor visual quality, including blur, underexposure, overexposure, and near-blank images. We do not apply a dedicated watermark filter. A human audit of 50K randomly sampled images from the final GPIC dataset identified 45 images with visual quality defects (0.09%) and 80 images with visible watermarks (0.16%).

(2) Safety filtering. We use a multi-stage safety pipeline combining Qwen3-VL-4B, which classifies images under a multi-category safety taxonomy including sexual, profane and violent content, and the FalconsAI NSFW detector at a low threshold of 0.30 to prioritize recall (Appendix D.1, Table D.1). An image is removed if either detector flags it. A human safety audit of 50K randomly sampled images from the final dataset identified 12 cases (0.024%) missed by both detectors, consisting of suggestive but non-explicit content. While no automated detection pipeline is perfect, this indicates a very low residual rate.

 ~

Q4: Caption quality and crop–caption consistency

We thank the reviewer for the question. We describe our captioning model selection and evaluation in Section 2.4 and Appendices F–G. We provide additional human audit and image–text alignment results below.

(1) Human validation of caption quality. We conducted a human audit of 50K randomly sampled image–caption pairs from the final GPIC dataset and identified 620 cases (1.24%) with captioning errors. These errors involve counting and spatial relationships in medium and long captions, while the overall scene descriptions remain accurate. We did not observe gross image–text mismatches.

(2) Image–text alignment and cropping. To measure image–text alignment, we use Qwen3-VL-Reranker-8B [G], a multimodal reranking model that estimates the probability (from 0 to 1) that an image matches a given text. We score each caption against its full image and against 256×256 center and random crops. Alignment scores are consistent across all three subsets, and cropping reduces alignment by 0.04–0.05 relative to full images.

DatasetFullCenter Crop 256Random Crop 256
GPIC-Nano (1M)0.9050.8660.852
GPIC-Val (200K)0.9060.8670.853
GPIC-Test (1M)0.9060.8670.853

We hope the rebuttal addresses all your concerns. We are happy to answer any additional questions.

 ~

References

[A] Peebles et al. Scalable diffusion models with transformers. ICCV 2023.

[B] Li et al. Back to basics: Let denoising generative models denoise. CVPR 2026.

[C] Li et al. Autoregressive image generation without vector quantization. NeurIPS 2024.

[D] Li et al. Return of unconditional generation. NeurIPS 2024.

[E] Zheng et al. Diffusion transformers with representation autoencoders. arXiv 2025.

[F] Yu et al. Representation alignment for generation. ICLR 2025.

[G] Li et al. Qwen3-VL-Embedding and Qwen3-VL-Reranker: A unified framework for state-of-the-art multimodal retrieval and ranking. arXiv 2026.

Reviewer_cZgJOpenReview ↗

Response to the rebuttal

After carefully read the rebuttal, I will keep my score considering that there seems little something new on the dataset distribution and the quality of the captions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author–Reviewer Discussion Summary

Dear AC and Reviewers,

As the author–reviewer discussion ends, we summarize reviewer concerns, the requested experiments, and clarifications provided in our rebuttal.

Contribution. GPIC's goal is to provide dataset and benchmarking infrastructure for academic research in visual generative modeling. GPIC provides 100M permissively licensed image–text pairs in fixed, centrally hosted shards, with 200K validation and 1M test examples, three nested training scales (GPIC-Nano, Lite, and Full), and a standardized FD-DINOv2 benchmarking protocol. No existing dataset or benchmark satisfies all four criteria: permissive, stable, large, and accessible (Table 1).

Reviewer concerns and responses:

  • Dataset and benchmark scope (cZgJ). GPIC is intended as a general-purpose dataset and benchmark for academic research in visual generation, not for training models that compete with frontier T2I generative models.
  • Licensing (XVku). Every record includes a license name, license URL, and attribution string; records with missing or malformed metadata were excluded. GPIC contains no ShareAlike, NonCommercial, or NoDerivatives images.
  • Human audits for image quality and safety (cZgJ, bE3o). Audits of 50K randomly sampled images from the final dataset show very low residual rates of quality defects (0.09%) and watermarks (0.16%), with no explicit content identified.
  • Human audit for caption quality (cZgJ, bE3o). Of 50K randomly sampled image–caption pairs, 1.24% contain captioning errors involving counting and spatial relationships in medium and long captions, while the overall scene descriptions remain accurate.
  • Image–text alignment and crop–caption consistency (cZgJ, bE3o). We provide Qwen3-VL-Reranker-8B alignment scores, which are consistent across GPIC-Nano, Val, and Test (0.905–0.906) and remain high on 256×256 center and random crops (0.852–0.867).
  • Other requested experiments and clarifications. A LlamaGen-XL (0.78B) baseline trained on GPIC-Full (cZgJ); stable model rankings across GPIC-Nano and GPIC-Full (XVku); deduplication and release of SSCD similarity scores (bE3o); captioning microbenchmark sampling (XVku); DINOv2 overlap in the benchmarking protocol (XVku); and future dataset updates (bE3o).

We hope these results and clarifications address the concerns raised by the reviewers.

Thank you for your time and consideration.

Best,
Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.