I'm broadly interested in generative AI, and specifically in generative vision models: their controllability, personalization, and efficiency. I want to build AI that closes the gap between what creative minds imagine and what they can put on screen.
Feel free to reach out. I'm open to collaborations!
*Currently looking for PhD positions — happy to chat.
Side quests:
MS Student in AIML
TU Darmstadt
News
Jun 2026June 2026
One co-first author paper accepted to ECCV 2026! 🎉 Looking forward to presenting Obliviate -- see you in Malmö! 🇸🇪
VETO is an anti-edit cloak that protects images from unauthorized editing by frontier AI models, disrupting how they read reference images through joint attention. We also introduce VetoBench, spanning closed-frame edits that modify the original scene and open-frame edits that recontextualize a subject in a newly generated one.
Selected
ECCV · 2026
Obliviate: Erasing Concepts from Autoregressive Image Generation Models
Obliviate erases concepts from autoregressive image generators by aligning conditional and pseudo-unconditional branches on shared visual prefixes, then applying KL supervision over full generation trajectories. Evaluated on Liquid, Emu3-Gen, and Janus-Pro, it removes targets like nudity or brands without collapsing model utility.
Under Review · 2026
Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models
We show that unified autoregressive models can be backdoored so a subtle trigger poisons both image and text outputs. ToBAC is the first such attack, effective via both data poisoning and model-level injection.
Under Review · 2026
The Poisoned Conversation: Privacy-Leaking Watermarks in Unified Multimodal Models
A dual-branch deep network detects diabetic retinopathy and grades its stage from a single fundus image. On APTOS 2019 it reaches 98.5% binary accuracy and strong stage-grading performance.
