SEED_balanced is the public balanced release of SEED, a benchmark for provenance tracing in sequential deepfake facial edits. Unlike conventional deepfake datasets that focus on single-step manipulations or binary real/fake detection, SEED models multi-step diffusion-based facial editing trajectories and supports three complementary tasks: Authenticity Analysis, Editing Trace Analysis, and Spatial Evidence Analysis. The full SEED benchmark contains 91,526 images… See the full description on the dataset page:
https://huggingface.co/datasets/Mengieong/SEED_balanced.
