WildGen

Data indices and tools for AI-generated image detection.

Associated publication: The Zero-Shot Illusion in the Wild: Diagnosing Boundary Failures in AIGC Detection.

Repository contents

Path	Contents
data/ai_index_public.csv	AI-generated image identifiers, source URLs, model metadata, labels, and split assignments.
data/real_experiment.csv	Real-image identifiers, source dataset metadata, image dimensions, labels, and split assignments.
metadata/image_sha256.csv	SHA-256 checksums of the original image files, keyed by image type and sample ID.
metadata/ai_source_licenses.csv	Source-level license names, upstream references, and evidence status for AI sources.
DATA_SOURCES.md	Real-data source references and instructions for checking applicable terms.
LEGACY_REFERENCES.md	Official repository locations for 16 baseline methods.
code/	Training scripts, evaluation adapters, data utilities, launchers, and dependency lists.

Data indices

Manifest	Total records	Training records	Validation records
AI images	342,000	307,800	34,200
Real images	347,207	339,207	8,000

The CSV split column uses train for training and test for validation. The Hugging Face configurations ai_index and real_index expose the manifests separately, each under a Hub split named index.

These files contain metadata and identifiers, not image binaries. Obtain images from their original sources and resolve their identifiers to local image paths before running the scripts. AI source URLs identify Civitai image pages.

Image checksums

metadata/image_sha256.csv contains 689,207 records with the columns image_type, sample_id, split, and sha256. Use image_type (ai or real) to select the corresponding index, then join on sample_id.

Each SHA-256 value identifies the original image file bytes, before decoding, resizing, or recompression. Compute SHA-256 on a downloaded file to check byte identity. The checksum is not a hash of its URL, sample ID, or decoded pixels. A re-encoded or modified image can have a different checksum even if it looks similar. The list contains no image binaries or local filesystem paths.

Content filtering and sensitive-content warning

For the Civitai-derived images, content filtering first excluded images flagged as potentially NSFW by Civitai. A private NSFW classifier was then applied as an additional screening stage to exclude images it classified as NSFW. The private classifier is not distributed with this repository.

Sensitive-content warning: Automated filtering is imperfect. Referenced images may still contain sexual, violent, disturbing, or otherwise sensitive material. The filtering process does not guarantee that every referenced image is suitable for all audiences. Review source content before displaying it or using it in applications with stricter content requirements.

Code contents

Paths below are relative to code/.

Training

File	Purpose
train_cnndetect.py	Fine-tune a ResNet-50 for binary real/AI classification.
train_univfd.py	Train a binary linear classifier on frozen CLIP features.
train_cnndetect_multiclass.py	Train a ResNet-50 for image-source classification.
train_univfd_multiclass.py	Train a source classifier on frozen CLIP features.

Data utilities

File	Purpose
scripts/data/csv_index_dataset.py	Load images from CSV indices and construct binary or source labels.
scripts/data/build_training_splits.py	Build seed-specific training subsets and fixed validation manifests.
scripts/data/build_transfer_manifests.py	Build GenImage and ForenSynths evaluation and joint-training manifests.

Evaluation

File	Purpose
scripts/eval/evaluate_pretrained_detectors.py	Evaluate pretrained training-based detectors.
scripts/eval/evaluate_dire.py	Run DIRE reconstruction and classification, and aggregate evaluation shards.
scripts/eval/evaluate_training_free.py	Evaluate RIGID, AEROBLADE, WaRPAD, and ManifoldBias.
scripts/eval/evaluate_real_source_fpr.py	Compute false-positive rates by real-image source.
scripts/eval/evaluate_binary_checkpoints.py	Evaluate final binary checkpoints on the fixed validation set.
scripts/eval/evaluate_feature_separability.py	Run frozen-feature binary probes and source-classification kNN diagnostics.
scripts/eval/evaluate_real_source_classifier.py	Train and evaluate a classifier for real-image sources.
scripts/eval/evaluate_robustness.py	Evaluate clean images and JPEG, resizing, and combined perturbations.

Launchers and family holdout

File	Purpose
scripts/launch/run_training_grid.py	Plan or execute binary, source-classification, and real/AI-ratio training grids.
scripts/launch/train_joint_adaptation.py	Run zero-shot evaluation or joint adaptation with labeled target data.
scripts/launch/run_transfer_evaluations.py	Launch cross-dataset evaluation, adaptation, and WildGen retention evaluation.
scripts/family_holdout/build_family_holdout_splits.py	Build training and evaluation splits using coarse generator-family labels.
scripts/family_holdout/train_family_probe.py	Extract and cache CLIP features, then train and evaluate a binary probe.

Supporting modules and dependencies

File	Purpose
legacy/evaluation/evaluate_legacy.py	Baseline model-loading and inference adapters.
legacy/evaluation/evaluate_real_only.py	Model-loading and inference adapters used for real-image evaluation.
legacy/evaluation/eval_utils.py	Shared baseline dataset and metric utilities.
legacy/__init__.py	Package initialization.
legacy/evaluation/__init__.py	Evaluation-package initialization.
env/pip_requirements.txt	Recorded training-environment package versions.
legacy/requirements.txt	Baseline evaluation dependencies.

Local inputs

Run executable scripts from code/ and use --help to inspect their arguments. Install the required baseline repositories and weights using LEGACY_REFERENCES.md. Third-party repositories and pretrained weights are not bundled.

The training-split builder uses four local CSV files: ai_train_pool.csv, ai_validation.csv, real_train_pool.csv, and real_validation.csv. Create them from the two public manifests while preserving identifiers and split assignments, and add resolved image_path values.

The AI interface uses sample_id, source_model_id (the public label), source_model_name (the public model_name), and image_path. The real interface uses sample_id, real_source_id, real_source_label, source_dataset, and image_path. The mapping YAML maps source names to integer class IDs, not Civitai platform IDs.

The family-holdout builder accepts the AI manifest's base_model and label columns, together with resolved image_path or relative_path values in a local copy of the index.

Third-party rights

The repository license does not grant rights to referenced third-party images, datasets, model weights, or upstream code. Those materials remain subject to their respective licenses and terms. Source URLs and identifiers are provided for provenance and do not constitute permission to reuse the referenced content.

Source-level AI license records are provided in metadata/ai_source_licenses.csv. The source_id column joins to label in the AI index. These are recorded model/resource terms, not blanket licenses for generated images. model_page indicates page-derived evidence; not_found and missing_model_id indicate that page evidence was unavailable in the source record. Blank fields do not imply unrestricted use. Consult DATA_SOURCES.md and the original sources for image-level and dataset-level terms.