Vishal Asnani / Research Scientist
Building trust
in this AI-generated era.
I’m a Research Scientist at Adobe Research in San Jose, working on trustworthy generative AI and content authenticity. My current work focuses on video and audio watermarking: building durable provenance signals that help verify digital content and preserve authenticity information as media is edited and shared.
At Adobe, I work on video and audio watermarking, with an emphasis on robustness and practical media provenance. My other research includes adaptable image watermarking, conditional-access watermarking, and concept-level attribution for generative models. I collaborate with John Collomosse and Shruti Agarwal, and mentor research interns on trustworthy media. Before joining full-time in January 2025, I interned at Adobe in 2023 and 2024, developing the research behind ProMark and CustomMark.
I earned my Ph.D. in Computer Science and Engineering at Michigan State University, advised by Prof. Xiaoming Liu. My doctoral research established the foundations of my work in proactive media forensics: embedding purposeful signals before manipulation to support detection and localization, and reverse engineering generative models from their outputs. That research continues to inform my work on traceable and verifiable media at Adobe.
Publications · CV · Talks & tutorials · Get in touch
Latest news
-
Our paper FiLM-CAM: Keyed Feature Modulation for Conditional-Access Watermarking has been accepted to the NeurIPS 2026 Main Track as a poster. Publication details will be shared soon.
Our paper Adaptable Watermarking: Transforming Image Watermarking Models via Lightweight Adapters has been accepted to WACV 2027. Publication details will be shared soon.
-
Signpost Watermarking, our work on joint optimization for visual watermark coexistence, was accepted to the ECCV 2026 LifegenIP Workshop.
-
FlowMark, our work on mask-guided video watermarking, is available as a preprint.
-
Co-organized Foundations and Frontiers of Watermarking at CVPR 2026, covering methods, multimodal extensions, robustness, and authenticity frameworks.
Our deepfake benchmark audit is available as a preprint.
TokenTrace and our edit-friendly hard prompt inversion work appeared at CVPR; PiVoT appeared at CVPR Workshops.
-
Our survey of proactive schemes appeared in the April issue of IJCV.
-
NRGMark appeared at WACV 2026, exploring localized watermarking for energy transparency in images.
-
Your Text Encoder Can Be An Object-Level Watermarking Controller appeared at ICCV, and CustomMark appeared at the ICCV APAI Workshop.
-
Joined Adobe Research as a Research Scientist.
-
Completed my Ph.D. at Michigan State University.
Our work on hyperparameter dependencies for model parsing appeared at NeurIPS 2024.
-
Shared the first version of our survey of proactive schemes, bringing together applications of purposeful signals in computer vision. The survey is now published in IJCV.
-
ProMark appeared at CVPR 2024, introducing proactive diffusion watermarking for causal attribution.
-
Returned to Adobe Research as a Research Scientist Intern, continuing research on generative AI attribution.
-
Shared the news that ProMark was accepted to CVPR 2024, following research during my 2023 Adobe internship.
-
Our Reverse Engineering of Deceptions survey was published in Foundations and Trends in Privacy and Security.
-
Our reverse engineering of generative models paper appeared in the December issue of TPAMI, exploring how generated images reveal model hyperparameters.
-
PrObeD, our proactive wrapper for object detection, was accepted to NeurIPS 2023.
-
Our work on reverse engineering generative models and inferring their hyperparameters was accepted to TPAMI.
-
MaLP, our proactive approach to manipulation localization, appeared at CVPR 2023.
Co-organized the Reverse Engineering of Deceptions tutorial at CVPR 2023.
-
Joined Adobe Research as a Research Scientist Intern, collaborating with John Collomosse and Shruti Agarwal on proactive attribution for generative AI.
-
Shared the news that MaLP: Manipulation Localization Using a Proactive Scheme was accepted to CVPR 2023.
-
Gave an in-person talk at Scale AI on reverse engineering generative models.
-
Gave a virtual talk at Scale AI on proactive image manipulation detection.
-
Proactive Image Manipulation Detection appeared at CVPR 2022, exploring learnable templates for detecting manipulated images.
-
Proactive Image Manipulation Detection was accepted to CVPR 2022.
Research interests
My current work at Adobe Research centers on video and audio watermarking for trustworthy media. My broader research interests build on my Ph.D. work in proactive manipulation detection, localization, and generative-model parsing.
Content authenticity
Signals that help media tell its story.
Robust image, video, and audio watermarking for traceable media, including durable provenance signals, localized metadata, and visual watermark coexistence. My current interests also include lightweight watermarking adapters and conditional-access watermarking.
Generative AI attribution
Tracing the ideas behind the image.
At Adobe, I explore how to trace the creative influences behind generated images through watermarked training data, diffusion-model customization, and multi-concept token recovery.
Media forensics & evaluation
Looking beyond what meets the eye.
Building on my Ph.D. research in proactive manipulation detection, localization, and model parsing, I investigate how to verify manipulated media, infer properties of generative models, and test what deepfake benchmarks really measure.
Illustrative research-area overviews, not experimental results.
Selected publications
Four perspectives on trustworthy media: watermarking video, tracing creative influences, localizing manipulation, and reverse engineering generative models.
FlowMark
FlowMark: Mask-Guided Video Watermarking
- The problem
- How can we add provenance signals to video, not just individual images?
- Our approach
- Explore mask-guided watermarking for video.
- Why it matters
- Extends the study of content provenance to moving imagery.
TokenTrace
TokenTrace: Multi-Concept Attribution through Watermarked Token Recovery
- The problem
- A generated image can combine several concepts. Which ones influenced it?
- Our approach
- Recover watermarked tokens to attribute generated images to multiple concepts.
- Why it matters
- Makes the mix of influences in generated content more traceable.
MaLP
MaLP: Manipulation Localization Using a Proactive Scheme
- The problem
- Knowing an image was manipulated is not enough. Where was it changed?
- Our approach
- Use a proactive template and local-global features to localize image manipulation.
- Why it matters
- Moves beyond a real-or-fake label to identifying manipulated regions.
Reverse engineering
Reverse Engineering of Generative Models: Inferring Model Hyperparameters from Generated Images
- The problem
- What can generated images reveal about the model that created them?
- Our approach
- Infer generative-model architecture and loss-function attributes from generated images.
- Why it matters
- Investigates the origins of synthetic media beyond detecting that it is generated.
Community and collaboration
I mentor research interns, co-organize tutorials, and review for venues including CVPR, ICCV, ECCV, TPAMI, IJCV, ICLR, AAAI, ACM Multimedia, and WACV. I also co-organized the CVPR 2023 tutorial Reverse Engineering of Deceptions and presented my research at Scale AI.
I welcome collaborations on video and audio watermarking, media provenance, content authenticity, generative AI attribution, and proactive media forensics. Reach me at vishal.8.asnani@gmail.com.
My collaborators have included Xiaoming Liu, Tal Hassner, John Collomosse, Shruti Agarwal, Xi Yin, and Sijia Liu.
Selected media coverage
Our research on reverse engineering generative models was covered by CNBC, The Verge, VentureBeat, The Wall Street Journal, and MSU Today.
Outside research
I enjoy long drives, jigsaw puzzles, reading, football, badminton, and table tennis.