The release provides the query images, a pre-generated reference pool, query-conditioned resynthesis references, and the exact reference-selection manifests used in the experiments, enabling reproducible evaluation of training-free synthetic image attribution.

Dataset Overview

BC-Attr-6 covers 10 text-to-image generators and six semantic categories.

For each generator-category pair, 200 images are included in total:

20 query images;
180 pre-generated reference images.

This gives:

10 generators;
6 semantic categories;
200 images per generator-category pair;
1,200 query images;
10,800 pre-generated reference images;
12,000 query and pre-generated reference images in total.

The pre-generated reference pool was constructed independently of the evaluation queries. The query images were generated later for attribution evaluation.

In addition, the release contains 120,000 query-conditioned resynthesis references, generated to provide references with stronger semantic correspondence to the query content.

Generators

The released generator labels are:

Flux2
Freepik
Lumina
Photon
Pixart
Playground
SD3
SD35
SDXL
Tencent_Hunyuan

Semantic Categories

The six semantic categories are:

animals
buildings
faces
panoramas
satellite_views
vehicles

Reference Regimes

The release supports three reference regimes.

1. Semantically Aligned References

References are selected from the pre-generated reference pool and restricted to the same semantic category as the query.

2. Arbitrary References

References are selected from the same pre-generated reference pool without requiring semantic alignment with the query.

3. Resynthesis References

References are generated conditionally on semantic information derived from the query.

These images are therefore not part of the pre-generated pool. They form a separate set of query-conditioned references.

The distinction can be summarized as:

query images
    │
    ├── pre-generated reference pool
    │       ├── arbitrary selection
    │       └── semantic-aligned selection
    │
    └── query-conditioned resynthesis
            └── resynthesis references

Repository Structure

BC-Attr-6/
├── README.md
│
├── query/
│   ├── animals/
│   │   ├── Flux2/
│   │   ├── Freepik/
│   │   └── ...
│   ├── buildings/
│   ├── faces/
│   ├── panoramas/
│   ├── satellite_views/
│   └── vehicles/
│
├── references/
│   ├── pre_generated/
│   │   ├── animals/
│   │   │   ├── Flux2/
│   │   │   ├── Freepik/
│   │   │   └── ...
│   │   ├── buildings/
│   │   ├── faces/
│   │   ├── panoramas/
│   │   ├── satellite_views/
│   │   └── vehicles/
│   │
│   └── resynthesis/
│       └── shards/
│           ├── shard-00000.tar
│           ├── shard-00001.tar
│           ├── ...
│           └── shard-00119.tar
│
├── metadata/
│   ├── images.csv
│   ├── query_prompt_content.csv
│   └── resynthesis_shards.csv
│
└── protocols/
    ├── semantic_aligned_reference.csv
    ├── arbitrary_reference.csv
    └── resynthesis_reference.csv

Query Images

The query/ directory contains 1,200 attribution queries.

Images are organized as:

query/<semantic>/<generator>/img_<id>.png

For example:

query/animals/Flux2/img_95.png
query/buildings/SDXL/img_42.png
query/vehicles/Tencent_Hunyuan/img_173.png

Each semantic-category–generator combination contains 20 query images.

Pre-generated Reference Pool

The directory:

references/pre_generated/

contains 10,800 pre-generated reference images.

Images are organized as:

references/pre_generated/<semantic>/<generator>/img_<id>.png

Each semantic-category–generator combination contains 180 reference-pool images.

The pool was constructed independently of the evaluation queries. The reference-selection protocols subsequently select images from this pool according to the arbitrary or semantically aligned regime.

The released protocols collectively use 7,800 unique images from the 10,800-image pool. The remaining 3,000 images are retained as part of the complete pre-generated reference pool; they are not stored in a separate directory.

Original generation prompts for pre-generated references are not required by the released attribution protocols.

Image Metadata

metadata/images.csv describes all 12,000 query and pre-generated reference images.

Columns:

Field	Description
`image_path`	Repository-relative image path
`role`	`query` or `pre_generated_reference`
`generator`	Source generator
`semantic`	Semantic category
`img_id`	Image identifier within the generator-category pair

Example:

image_path,role,generator,semantic,img_id
query/animals/Flux2/img_95.png,query,Flux2,animals,95
references/pre_generated/animals/Flux2/img_80.png,pre_generated_reference,Flux2,animals,80

The metadata contains:

query                     1,200
pre_generated_reference  10,800
total                    12,000

Query Content Prompts

metadata/query_prompt_content.csv contains the image-content descriptions used for query-conditioned resynthesis.

Columns:

Field	Description
`query_image_path`	Repository-relative query path
`generator`	Ground-truth source generator
`semantic`	Semantic category
`img_id`	Query image identifier
`prompt_content`	Content description used for resynthesis

Example:

query_image_path,generator,semantic,img_id,prompt_content
query/animals/Flux2/img_95.png,Flux2,animals,95,"..."

The file contains 1,200 rows, providing one content description for every released query image.

The content descriptions were obtained using LLaVA-1.5-7B and correspond to the semantic content prompts used for the reported resynthesis experiments.

Only prompt_content, which is used in the reported experiments, is included in the public release.

Experimental Reference Protocols

The exact reference realizations used for the experiments are provided under:

protocols/

These manifests should be treated as the authoritative experimental protocols.

Semantically Aligned References

File:

protocols/semantic_aligned_reference.csv

For every query and candidate generator, reference images are selected from:

references/pre_generated/

while requiring the reference image to belong to the same semantic category as the query.

The manifest contains 100 ordered reference candidates per query-candidate-generator pair, for a total of:

1,200 queries
× 10 candidate generators
× 100 references
= 1,200,000 rows

Columns:

Field	Description
`query_image_path`	Query image path
`semantic`	Query semantic category
`query_generator`	Ground-truth source generator of the query
`img_id`	Query image identifier
`reference_generator`	Candidate source generator
`sample_id`	Ordered reference index
`reference_image_path`	Selected pre-generated reference image

Arbitrary References

File:

protocols/arbitrary_reference.csv

For every query and candidate generator, reference images are selected from the pre-generated reference pool without requiring semantic alignment with the query.

The manifest likewise contains:

1,200,000 rows

with 100 ordered reference candidates for each query-candidate-generator pair.

Its schema is identical to the semantically aligned protocol:

Field	Description
`query_image_path`	Query image path
`semantic`	Query semantic category
`query_generator`	Ground-truth source generator of the query
`img_id`	Query image identifier
`reference_generator`	Candidate source generator
`sample_id`	Ordered reference index
`reference_image_path`	Selected pre-generated reference image

Reference Budget

The pre-generated reference manifests preserve the exact ordering used during evaluation through sample_id.

For a reference budget (M), retain the first (M) references:

subset = protocol[protocol["sample_id"] < M]

For example:

M = 1   -> sample_id 0
M = 5   -> sample_id 0 ... 4
M = 10  -> sample_id 0 ... 9
M = 20  -> sample_id 0 ... 19

Because all budgets are prefixes of the same ordered list, the released manifests preserve the exact reference realization instead of requiring references to be sampled again.

Query-Conditioned Resynthesis References

The repository also provides the references generated conditionally on query-side semantic information.

For each query:

10 candidate generators are considered;
10 references are synthesized per candidate generator;
1,200 queries are included.

Therefore:

1,200 queries
× 10 candidate generators
× 10 samples
= 120,000 resynthesis references

The authoritative manifest is:

protocols/resynthesis_reference.csv

Resynthesis Manifest

Columns:

Field	Description
`query_image_path`	Query image path
`semantic`	Query semantic category
`query_generator`	Ground-truth source generator
`img_id`	Query image identifier
`prompt_type`	Resynthesis prompt regime
`prompt_text`	Prompt used for resynthesis
`reference_generator`	Candidate generator used to synthesize the reference
`sample_id`	Resynthesis sample index
`reference_image_path`	Logical repository-relative path of the resynthesis reference
`shard_path`	TAR shard physically containing the image
`member_path`	Image path inside the TAR shard
`sha256`	SHA256 checksum of the PNG bytes
`size_bytes`	Original PNG file size

For every query-candidate-generator pair:

sample_id = 0, 1, ..., 9

The generation seed corresponds to the sample index used during resynthesis.

Resynthesis Storage

The 120,000 resynthesis references are stored in 120 TAR shards, with 1,000 PNG files per shard:

references/resynthesis/shards/shard-00000.tar
references/resynthesis/shards/shard-00001.tar
...
references/resynthesis/shards/shard-00119.tar

The images are stored without image recompression or conversion.

No resizing, format conversion, or PNG re-encoding was performed during sharding. The original PNG bytes are preserved inside the TAR archives.

Resynthesis Shard Index

metadata/resynthesis_shards.csv maps every logical resynthesis reference to its physical TAR location.

Columns:

Field	Description
`reference_image_path`	Logical repository-relative reference path
`shard_path`	TAR shard path
`member_path`	Path inside the TAR archive
`sha256`	SHA256 checksum
`size_bytes`	PNG size in bytes

Example logical paths have the form:

references/resynthesis/<semantic>/<query_generator>/prompt_content/<candidate_generator>/img_<id>_<sample_id>.png

Reading a Resynthesis Reference

A resynthesis reference can be accessed through its shard_path and member_path.

Example:

import io
import tarfile

from PIL import Image

shard_path = "references/resynthesis/shards/shard-00000.tar"
member_path = "animals/Flux2/prompt_content/SDXL/img_95_3.png"

with tarfile.open(shard_path, "r") as tar:
    f = tar.extractfile(member_path)
    image = Image.open(io.BytesIO(f.read())).convert("RGB")

print(image.size)

Alternatively, an entire shard can be extracted with:

mkdir -p references/resynthesis

tar -xf references/resynthesis/shards/shard-00000.tar \
    -C references/resynthesis/

Because the TAR member paths preserve the logical directory structure, extraction reconstructs paths of the form:

references/resynthesis/<semantic>/<query_generator>/prompt_content/<candidate_generator>/img_<id>_<sample_id>.png

Integrity Verification

The release was subjected to consistency and byte-level integrity checks.

The verified release contains:

Query images:                         1,200
Pre-generated reference images:      10,800
Query + pre-generated images:        12,000

Resynthesis references:             120,000
Resynthesis TAR shards:                 120
Images per shard:                      1,000
Total TAR members:                   120,000

The query and pre-generated reference corpus satisfies:

60 generator-category combinations
× 200 images per combination
= 12,000 images

with:

20 query images per combination
180 pre-generated references per combination

For every resynthesis reference:

SHA256 was computed from the original PNG;
the PNG bytes were written into the corresponding TAR archive;
the TAR member was read back from the completed archive;
its byte size was verified;
its SHA256 was recomputed;
the resulting hash was verified against the original PNG hash.

No image decoding, resizing, conversion, or recompression was involved in the sharding process.

Reproducing the Experimental Protocol

Pre-generated references

Choose either:

protocols/semantic_aligned_reference.csv

or:

protocols/arbitrary_reference.csv

Then select the desired reference budget (M):

df = df[df["sample_id"] < M]

The resulting rows directly identify the corresponding query and pre-generated reference images.

Resynthesis references

Use:

protocols/resynthesis_reference.csv

to identify:

the query;
candidate source generator;
resynthesis prompt;
sample index;
logical reference path;
physical shard;
TAR member path.

The released manifests preserve the exact reference realizations used during evaluation.

Dataset Statistics

Query and Pre-generated Reference Corpus

Property	Value
Generators	10
Semantic categories	6
Generator-category combinations	60
Images per generator-category pair	200
Query images per pair	20
Pre-generated references per pair	180
Query images	1,200
Pre-generated reference pool	10,800
Total	12,000

Pre-generated Reference Protocols

Protocol	Rows	Maximum references per query-source pair
Semantically aligned	1,200,000	100
Arbitrary	1,200,000	100

Across the two released protocols, 7,800 unique pre-generated reference images are selected from the complete 10,800-image pool.

The remaining 3,000 images remain part of the released reference pool but are not selected by these specific protocol realizations.

Query-Conditioned Resynthesis

Property	Value
Queries	1,200
Candidate generators	10
Samples per query-generator pair	10
Total resynthesis references	120,000
TAR shards	120

Intended Use

BC-Attr-6 is intended for academic research on topics including:

synthetic image source attribution;
AI-generated image forensics;
reference-based attribution;
training-free attribution;
representation analysis;
reference selection;
robustness and generalization of synthetic image attribution methods.

The released protocol manifests should be used when reproducing the experiments reported in the associated paper.

Code

The implementation of the experimental pipeline, including:

prompt inversion;
query-conditioned resynthesis;
reference construction;
reference selection;
CLIP feature extraction;
DINOv2 feature extraction;
attribution evaluation;

is released separately through the associated code repository.

The GitHub link will be added here when the public repository is available.

Citation

If you use BC-Attr-6 or its associated experimental protocols, please cite:

@article{li2026representation,
  title={Representation and Reference Selection in Training-Free Synthetic Image Attribution},
  author={Li, Meiling and Bongini, Pietro and Tondi, Benedetta and Barni, Mauro},
  journal={arXiv preprint arXiv:2607.12052},
  year={2026}
}

Usage and Licensing

BC-Attr-6 is released for research purposes.

The dataset contains synthetic images produced by multiple image-generation systems. Users are responsible for complying with the applicable terms and licenses of the corresponding source models and services.

Final dataset licensing and access conditions will be specified before public release.

Contact

For questions regarding the dataset or the experimental protocols, please contact the authors of the associated paper.

Downloads last month: 69

Total file size:

186 GB

Paper for Meilinger00/BC-Attr-6

Representation and Reference Selection in Training-Free Synthetic Image Attribution

Paper • 2607.12052 • Published Jul 13