HEAR — CLIP-L/14 checkpoint

Checkpoint for HEAR: Hierarchical Evidence Accumulation with Coherence Refinement for Deepfake Face Video Detection.

This is the original research checkpoint, trained using the FaceForensics++ c23 setup. It is supplied as a PyTorch state dictionary (hear_clip_l14.pth), not as a Transformers from_pretrained model. Load it using the HEAR repository's training/test.py and the included clip_inference.yaml configuration. Labels are 0 for real and 1 for fake.

The cleaned HEAR loader maps legacy appearance-head parameter names and discards the removed auxiliary temporal-difference branch weights. Predictions use the appearance evidence path. This checkpoint is for CLIP-L/14, not DINOv3.

The training, testing, and checkpoint-selection framework and major portions of the implementation are adapted from DeepfakeBench. Evaluation datasets and access instructions are provided by DeepfakeBench and DF40.

A model score is a prediction, not proof that a video is authentic or manipulated. Performance can change across datasets, manipulation methods, and preprocessing pipelines. No new benchmark results are asserted by this model card.