CoVAtt-Benchmark

A large-scale benchmark for generated-image attribution: given an image that was produced by some text-to-image model, decide which model produced it, and decide whether it came from a model the system has never seen before.

What this dataset is

This is the dataset used to train and evaluate CoVAtt (Content-Based Verification for Attribution of AI-Generated Images, BMVC 2026). CoVAtt is a Siamese network that takes a pair of images and predicts whether they came from the same generator. That single pairwise decision is then used two ways:

• Closed-set attribution — match a query image against reference images from each known generator and attribute it to the best-matching one.
• Open-set detection — flag a query image whose similarity to every known generator is too low, i.e. it came from a generator outside the known set.

The dataset itself is generator-agnostic, so it is equally usable for plain closed-set classification, synthetic-image detection, or any other attribution method.

Contents: 254,893 generated images from 13 text-to-image generators, spanning the field from 2021 to 2024 — early GAN/diffusion systems (GALIP, GLIDE, LDM), the Stable Diffusion line (SD1.4 through SD3.5), modern high-quality models (FLUX, Stable Cascade), and distilled few-step variants (SDXL-Turbo, SD3.5-Turbo, Hyper-SD). Total size is about 245 GB, packaged as WebDataset-style .tar shards of roughly 1 GB each.

Two design properties that make it an attribution benchmark

1. Content is held constant across generators. Every generator was prompted with the same fixed list of COCO captions. So for any given caption, there is a corresponding image from each of the 13 generators depicting the same described scene. The systematic difference between two generators' folders is therefore the generator itself, not the subject matter — which is what forces a model to key on generator fingerprint rather than on image content.

2. Every generator was run twice (v1 and v2). The two batches cover the same captions but are independent generation runs, so a caption's v1 and v2 images depict the same scene while differing in sampling noise. This yields same-generator image pairs that share no pixels, which is what lets a pairwise model learn "same generator" as distinct from "same prompt".

Quick start

Each generator is exposed as its own config, with v1 and v2 as splits. The dataset is large, so streaming is recommended:

from datasets import load_dataset

# One generator, one batch
ds = load_dataset("kenyag/CoVAtt-Benchmark", "SDXL", split="v1", streaming=True)
sample = next(iter(ds))
sample["png"]        # PIL image
sample["__key__"]    # sample_id, joins to metadata/samples.parquet

To recover the caption behind an image, or to work across generators, join on sample_id using the manifest:

import pandas as pd
from huggingface_hub import hf_hub_download

manifest = pd.read_parquet(hf_hub_download(
    "kenyag/CoVAtt-Benchmark", "metadata/samples.parquet", repo_type="dataset"))
manifest.head()   # sample_id, generator, batch, shard_file, member_name, caption, ...

Layout

CoVAtt-Benchmark/
│
├── README.md
├── LICENSE
├── CITATION.cff
│
├── metadata/
│   ├── samples.parquet           # per-image manifest: sample_id, generator,
│   │                               batch, shard_file, member_name, caption,
│   │                               original_filename, file_size_bytes
│   ├── captions.csv              # master caption list used as the
│   │                               generation prompts, with COCO filenames
│   ├── generators.csv            # per-generator checkpoint, LoRA/base info,
│   │                               license, per-batch image counts
│   ├── generation_settings.yaml  # default-settings note
│   └── licenses/                 # <Generator>.txt: that checkpoint's
│                                   license + what it says about outputs
│
└── data/
    ├── GLIDE/{v1,v2}/part-NNN.tar
    ├── GALIP/{v1,v2}/part-NNN.tar
    ├── LDM/{v1,v2}/part-NNN.tar
    ├── SD1.4/{v1,v2}/part-NNN.tar
    ├── SDXL/{v1,v2}/part-NNN.tar
    ├── SDXL-Turbo/{v1,v2}/part-NNN.tar
    ├── Cascade/{v1,v2}/part-NNN.tar
    ├── Hyper-SD/{v1,v2}/part-NNN.tar
    ├── SD3/{v1,v2}/part-NNN.tar
    ├── SD3.5/{v1,v2}/part-NNN.tar
    ├── SD3.5-Turbo/{v1,v2}/part-NNN.tar
    ├── FLUX/{v1,v2}/part-NNN.tar
    └── DALL-E/{v1,v2}/part-NNN.tar

Archive members are named by a zero-padded numeric sample_id, not by the caption text, because a few captions contain characters that are illegal in paths. The original caption and filename are preserved in metadata/samples.parquet.

Exact per-generator, per-batch image counts are in metadata/generators.csv. They are near-identical across generators but not exactly equal: the two generators run outside HuggingFace diffusers (GLIDE and GALIP) resolved a slightly larger portion of the caption list than the diffusers-based ones.

Generators and checkpoints

All generators were run with the default pipeline settings of the listed checkpoint (no custom guidance scale, step count, or scheduler overrides), except where noted. Full detail in metadata/generators.csv.

data/ folder	Model checkpoint	Checkpoint license	Notes
GLIDE	GLIDE (filtered): base.pt + upsample.pt, Dec 2021 release, via openai/glide-text2im	MIT	Only GLIDE checkpoint OpenAI ever released
GALIP	GALIP, COCO-pretrained (pre_coco.pth), via tobran/GALIP	Academic research use only (see note below)	GAN, not a diffusion model
LDM	CompVis/ldm-text2im-large-256	Apache-2.0
SD1.4	CompVis/stable-diffusion-v1-4	CreativeML OpenRAIL-M
SDXL	stabilityai/stable-diffusion-xl-base-1.0 (+ stabilityai/stable-diffusion-xl-refiner-1.0)	CreativeML Open RAIL++-M	Base + refiner
SDXL-Turbo	stabilityai/sdxl-turbo	Stability AI Community License Agreement	Distilled, few-step
Cascade	stabilityai/stable-cascade-prior + stabilityai/stable-cascade	Stability AI Non-Commercial Research Community License
Hyper-SD	ByteDance/Hyper-SD LoRA (Hyper-SDXL-1step-lora.safetensors) on stabilityai/stable-diffusion-xl-base-1.0	ByteDance Hyper-SD License	1-step distilled LoRA
SD3	stabilityai/stable-diffusion-3-medium-diffusers	Stability AI Community License Agreement
SD3.5	stabilityai/stable-diffusion-3.5-large	Stability AI Community License Agreement
SD3.5-Turbo	stabilityai/stable-diffusion-3.5-large-turbo	Stability AI Community License Agreement	Distilled, few-step
FLUX	black-forest-labs/FLUX.1-dev	FLUX.1 [dev] Non-Commercial License
DALL-E	fluently/Fluently-XL-v2 + LoRA ehristoforu/dalle-3-xl-v2	CreativeML OpenRAIL-M (LoRA) + Fluently Models License (base)	DALL-E-3-style LoRA on an SDXL finetune, not the OpenAI DALL-E 3 API — named to avoid misattribution

Full per-generator license text, and what each license says about generated outputs specifically, is in metadata/licenses/<Generator>.txt and metadata/generators.csv.

Real images

The captions come from COCO train2017 (https://cocodataset.org). The real COCO images themselves are not redistributed here — if your setup needs real images alongside the generated ones, download them from COCO directly under its own license and terms. metadata/captions.csv records the COCO filename each caption came from, so generated images can be aligned back to their real counterparts.

Limitations and intended use

• Captions come from a single source domain (COCO: everyday scenes, objects, people). Attribution performance measured here may not transfer unchanged to other prompt distributions such as portraits, art styles, or text-heavy images.
• All images are stored as originally generated, without post-processing. Robustness to JPEG recompression, resizing, or social-media pipelines is not represented in this data and has to be simulated separately.
• Generators were run at their default settings. Changing sampler, step count, or guidance scale can shift a generator's fingerprint, so a model trained only on this data may degrade on non-default sampling.
• This dataset is intended for research on attribution, provenance, and synthetic-image detection. It is not intended for training generative models.

License

Two layers apply:

• The dataset compilation (folder/shard structure, metadata/samples.parquet, captions.csv, generators.csv, generation_settings.yaml, README, citation) is released under CC BY-NC-SA 4.0. See LICENSE.
• Each generated image additionally carries the license of the checkpoint that produced it — see the Checkpoint license column above and metadata/licenses/<Generator>.txt for the full text and what that license says about generated outputs specifically. Most of these checkpoints disclaim any ownership of or restriction on their outputs, even where the model weights themselves are non-commercial (FLUX.1-dev, Stable Cascade, SDXL-Turbo). GALIP is the exception: its license is academic-research-only, and that restriction is treated as applying to its images in this release too.

Citation

@inproceedings{Yaggel_2026_BMVC,
  author    = {Ken Yaggel and Edita Grolman and Hiroo Saito and Yuto Yamaji and Misaki Komatsu and Yoshikazu Hanatani and Asaf Shabtai and Yuval Elovici},
  title     = {CoVAtt - Content-Based Verification for Attribution of AI-Generated Images},
  booktitle = {British Machine Vision Conference 2026, {BMVC} 2026},
  publisher = {BMVA},
  year      = {2026},
  note      = {Accepted; proceedings forthcoming}
}

See also CITATION.cff.