Participants will be provided with a dataset released exclusively for this challenge. The dataset
itself is composed of 10 open-source text-to-image models (all available on Hugging Face).
In particular, it is obtained by taking the prompt list of the WILD dataset 2, and using the
prompts to generate new images with the sources. The dataset includes:
• a labeled training set;
• an unlabeled test set → ground-truth labels of the test set will remain private.
The dataset may not be redistributed or used outside the scope of this challenge.
The dataset is composed of 10,000 images (each source contributed 1,000 images). The training set contains 7,000 images, while the test set contains 3,000 images. All the test set images have undergone one to three post-processing operation(s) randomly drawn from the following set: 'jpeg compression', 'webp compression' 'random central crop', 'resizing', 'small rotation with aspect preserving crop', 'contrast/brightness adjustment', 'Gaussian blur', 'grayscale conversion', 'AI  super-resolution using pre-trained models', 'JPEG AI compression'.

The sources are: AuraFlow, Freepik, Lumina, Photon, Pixart(sigma), Playground v2.5, StableDiffusion3, StableDiffusion3.5, StableDiffusionXL-Turbo, Tencent Hunyuan.

The objective is to develop a synthetic image source attribution method that can attribute test set images with better performance with respect to competitors. In particular, since we have a closed set of 10 sources, the problem can be seen as a 10-class classification task.

Once a model is trained, it can be evaluated on the test set by submitting a .csv file. Half of the test set will be used to provide an evaluation for the public leaderboard (fast feedback) while half of the test set will be reserved for the final evaluation that will produce the private leaderboard. The final leaderboard will be published after the challenge completion.

Please note that the submission files should be structured as detailed in the following.