Maty Bohacek

Stanford & Google DeepMind — San Francisco, CA

I am a student at Stanford University and student researcher at Google DeepMind. Advised by Prof. Hany Farid and Prof. Maneesh Agrawala, I work at the intersection of AI, computer vision, and media forensics. My goal is to build AI systems that are inherently trustworthy and interpretable.

News

(September 2025)

  • Received the Czech Senate Medal (a civilian honor) from Senate Speaker Milos Vystrcil in Prague.

(August 2025)

  • Our IJCAI-W paper ‘Human Action CLIPs’ won the best paper award (conference held in Montreal, Canada).

  • See full archive in my journal here.

Upcoming

(October 2025)

  • I will deliver remarks at the Planeta AI conference.

  • I will deliver a talk at the INN conference.

  • Hosting a workshop and two workshop papers at ICCV in Honolulu, HI.

(November 2025)

  • One paper at ACM MM in Dublin, Ireland.

Highlighted Publications

* Equal contribution. Not necessarily chronological.

Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders

Bohacek M.*, Fel T.*, Agrawala M., Lubana E. Under Review.
Project Paper Code Weights — Method for identifying blindspots in T2I models: concepts that the model was trained on but can’t generate.

Human Action CLIPs: Detecting AI-generated Human Motion

Bohacek M. & Farid H. IJCAI-W 2025.
Paper Dataset — This paper proposes a method for distinguishing real and fake (T2V) video using multi-modal semantic embeddings, evaluated on DeepAction, a new dataset of real and AI-generated human motion.

The DeepSpeak Dataset

Barrington S., Bohacek M., and Farid H. ArXiv, abs/2408.05366.
Paper Dataset — This paper introduces DeepSpeak, a large-scale dataset of real and deepfake footage designed to support research on detecting state-of-the-art face-swap and lip-sync deepfakes.

Nepotistically Trained Generative-AI Models Collapse

Bohacek M. & Farid H. ICLR-W 2025.
Paper — This paper demonstrates how some generative AI models, when retrained on their own outputs, produce distorted images and struggle to recover even after retraining on real data.

GenAI Confessions: Black-box Membership Inference for Generative Image Models

Bohacek M. & Farid H. ICCV-W 2025.
Paper Dataset — A method to determine whether a generative AI model was trained on specific images.

Synthetic Human Action Video Data Generation with Pose Transfer

Knapp V. & Bohacek M. CVPR-W 2025.
Project Paper Code Data — We show that synthetic data of human motion improves performance of action classification and understanding.

For a complete list of my academic publications, please refer to this page or my Google Scholar profile.

Contact & Misc.

  • Email: maty (at) stanford (dot) edu

  • Resume (coming soon)