COCO-Attr

COCO-Attr is the COCO-based evaluation dataset used in:

Representation and Reference Selection in Training-Free Synthetic Image Attribution

Paper: https://arxiv.org/abs/2607.12052

The dataset contains images from 10 synthetic image generators and supports two reference-based source attribution protocols:

• Arbitrary references
• Resynthesis references

Generators

The dataset contains images from the following 10 generators:

1. Flux2
2. Freepik
3. Lumina
4. Photon
5. Pixart
6. Playground
7. SD3
8. SD35
9. SDXL
10. Tencent_Hunyuan

Dataset Statistics

Component	Count
Query images	1,000
Pre-generated reference images	39,429
Resynthesis reference images	100,000
Resynthesis TAR shards	100
Arbitrary-reference protocol rows	1,000,000
Resynthesis-reference protocol rows	100,000

The query set contains 100 images per generator.

The query images and pre-generated reference images are derived from disjoint COCO source splits. Only the 1,000 validation-side images actually used as queries in the reported experiments are included.

Repository Organization

Query images

query/

Contains the 1,000 evaluation queries, organized by generator.

Pre-generated references

references/pre_generated/

Contains the 39,429 images forming the arbitrary-reference pool, organized by generator.

Resynthesis references

references/resynthesis/shards/

Contains 100 TAR shards. Each shard stores exactly 1,000 resynthesized reference images.

Metadata

metadata/images.csv

Metadata for the query images and pre-generated reference images.

metadata/query_prompt_content.csv

Content descriptions extracted from the 1,000 query images and used for resynthesis.

metadata/resynthesis_shards.csv

Shard-level metadata, including shard path, number of files, size, and SHA256 checksum.

Protocols

protocols/arbitrary_reference.csv

Defines the arbitrary-reference assignments used in the experiments.

protocols/resynthesis_reference.csv

Defines the resynthesis-reference assignments and maps every resynthesized image to its TAR shard and member path.

Arbitrary-Reference Protocol

For each query and each candidate generator, the released protocol contains up to 100 pre-generated references.

The CSV preserves the exact reference assignments used in the experiments. A reference budget of (M) is obtained by selecting rows satisfying:

sample_id < M

For example, the reported experiment with (M=10) uses reference samples with sample_id from 0 to 9.

The released arbitrary-reference protocol corresponds to the experimental random_full setting.

Resynthesis-Reference Protocol

For each query and each candidate generator, 10 references are generated by resynthesizing the semantic content of the query with that candidate generator.

Therefore, the dataset contains:

1,000 queries ?? 10 candidate generators ?? 10 samples = 100,000 resynthesis references.

The resynthesis protocol provides:

• query image path
• ground-truth query generator
• candidate generator
• prompt
• sample ID
• logical image path
• TAR shard path
• TAR member path
• SHA256 checksum
• file size

Resynthesis Storage

The 100,000 resynthesis images are stored in 100 TAR shards rather than as individual repository files.

Each shard contains exactly 1,000 images.

metadata/resynthesis_shards.csv records the SHA256 checksum and size of every shard.

protocols/resynthesis_reference.csv additionally records the SHA256 checksum and file size of every individual resynthesis image.

Integrity Verification

Before release, the dataset was validated to ensure that:

• all 1,000 query images are present;
• all 39,429 pre-generated reference images are present;
• all 100,000 resynthesis images are represented;
• all 100 TAR shards contain exactly 1,000 files;
• all protocol paths are repository-relative;
• TAR members exactly match the protocol manifest;
• every resynthesis image was verified by SHA256 after being read back from its TAR shard;
• shard-level SHA256 checksums are recorded.

Citation

If you use COCO-Attr, please cite:

Representation and Reference Selection in Training-Free Synthetic Image Attribution

https://arxiv.org/abs/2607.12052

Related Dataset

BC-Attr-6 is the complementary benchmark used in the same work for evaluating arbitrary, semantically aligned, and resynthesis reference selection.