News
The lab moves to Colorado State University
SeeVi relocates from Lehigh to the Department of Computer Science at Colorado State University in Fort Collins, where the lab continues its work on provenance, media integrity, and visual perception.
Qiuyu Tang presented StyleProtect in Denver. The method perturbs only the cross-attention layers that carry style rather than the whole model, letting artists resist style mimicry from finetuned diffusion models at lower cost.
Chaitanya Roygaga’s paper asks whether vision-language models reproduce the snap judgements people make about faces, probing them with generated captions instead of fixed templates.
VALED — Visual Art, Generative AI, and the Legal/Ethical Dilemma — brings together researchers, artists, legal scholars, and industry practitioners to examine style mimicry, dataset auditing, provenance, and the copyright questions raised by text-to-image generation.
Led by Qiuyu Tang, the paper tests PhotoGuard, Mist and Glaze and finds they do not disrupt editing as intended. Adding protective noise often makes the resulting edit align more closely with the attacker’s prompt, not less.
With collaborators at Unicamp and Loyola University Chicago. Neural codecs such as JPEG-AI preserve both the meaning of scientific images and the forensic traces of tampering better than standard JPEG at similar bitrates.
Part of a line of work on why vision-language models handle affirmation well and negation poorly.