I am a Ph.D. student in the ECE Department at UT Austin, where I am fortunate to be advised by Prof. Diana Marculescu. I received my master's and bachelor's degrees from National Taiwan University.
My research focuses on generative AI for computer vision, with a particular interest in efficiency. If you find that we have shared research interests, please feel free to reach out by email; I am always open to potential collaborations.
Education
Ph.D. in ECE department at University of Texas at Austin, advised by Prof. Diana Marculescu, August 2023 — Present.
M.S. in ECE at National Taiwan University, advised by Prof. Lin-shan Lee, GPA: 4.26/4.30, August 2021 — July 2023.
B.S. in ECE at National Taiwan University, GPA: 4.03/4.30, August 2017 — June 2021.
Training-Free Latent Inter-frame Pruning with Attention Recovery
Dennis Menn, Yuedong Yang, Bokun Wang, et al.
Proposed a training-free framework that adapts pixel-space compression techniques to latent pruning for video generation. Achieved an average 1.53× throughput boost (19.3 FPS on an RTX 4090) and a 20% memory reduction, without compromising visual quality.
The top half row displays original images, while the bottom half row shows corrupted images with green lines denoting the selected region to inject artifacts.
Synthesizing Artifact Dataset for Pixel-level Detection
Dennis Menn, Feng Liang, and Diana Marculescu
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026, under review.
We propose a pipeline that automatically injects artifacts into clean, high-quality synthetic images on a predetermined region, thereby producing pixel-level annotations without manual labeling.
Comparison of an image exhibiting artifacts (left) and a natural-looking image (right), along with their corresponding Similarity Trajectories.
Similarity Trajectories: Linking Sampling Process to Artifacts in Diffusion-Generated Images
Dennis Menn, Feng Liang, Hung-Yueh Chiang, and Diana Marculescu
4th Workshop on Image/Video/Audio Quality in Computer Vision and Generative AI, IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2025
We found that the presence of artifacts in images is related to the sampling process of the diffusion model, characterized by the "Similarity Trajectory".
Investigating Human-Identifiable Features Hidden In Adversarial Perturbations
Dennis Menn, Tzu-hsun Feng, Sriram Vishwanath, and Hung-yi Lee
Manuscript
By summing adversarial perturbations from different models, human-identifiable features emerge. Showing that adversarial perturbations are in fact comprehensible to humans.
Can GPT Models Detect Artifacts in Images?
Dennis Menn, Alexandros Dimakis, and Diana Marculescu
Blog post
GPT-4/4o models are good at recognizing objects but struggle to identify obvious errors in images generated by diffusion models. One explanation is that they only recognize key features rather than understanding the complete structure of objects.
