The rapid advancement of talking-head deepfake generation fueled by advanced generative models has elevated the realism of synthetic videos to a level that poses substantial risks in domains such as media, politics, and finance. However, current benchmarks for deepfake talking-head detection fail to reflect this progress, relying on outdated generators and offering limited insight into model robustness and generalization. We introduce TalkingHeadBench, a new benchmark designed to address this gap, featuring talking-head videos from six modern generators, with an additional two emerging generators used exclusively for testing generalization. The dataset is built on an expert-led curation process that filters over 60% of samples to remove videos with noticeable artifacts, presenting a more difficult challenge for detectors. Our evaluation protocols are designed to measure generalization across identity and generator shifts. Benchmarking seven state-of-the-art detectors reveals that models with high accuracy on older datasets like FaceForensics++ show a significant performance drop on our curated data, particularly at strict false positive rates (e.g., TPR@FPR=0.1%). In addition, we identify a trend where detectors focus on background cues instead of facial features using Grad-CAM visualization. Our benchmark aims to accelerate research towards more robust and generalizable detection models in the face of rapidly evolving generative techniques. We release our benchmark and dataset with all data splits and protocols at https://anaxqx.github.io/talkingheadbench.github.io.

We present a novel usage of Transformers to make image classification interpretable. Unlike mainstream classifiers that wait until the last fully connected layer to incorporate class information to make predictions, we investigate a proactive approach, asking each class to search for itself in an image. We realize this idea via a Transformer encoder-decoder inspired by DEtection TRansformer (DETR). We learn ''class-specific'' queries (one for each class) as input to the decoder, enabling each class to localize its patterns in an image via cross-attention. We name our approach INterpretable TRansformer (INTR), which is fairly easy to implement and exhibits several compelling properties. We show that INTR intrinsically encourages each class to attend distinctively; the cross-attention weights thus provide a faithful interpretation of the prediction. Interestingly, via ''multi-head'' cross-attention, INTR could identify different ''attributes'' of a class, making it particularly suitable for fine-grained classification and analysis, which we demonstrate on eight datasets. Our code and pre-trained models are publicly accessible at the Imageomics Institute GitHub site: https://github.com/Imageomics/INTR.