InvAD: Inversion-based Reconstruction-Free Anomaly Detection with Diffusion Models
Shunsuke Sakai, Xiangteng He, Chunzhi Gu, Leonid Sigal, Tatsuhito Hasegawa
CVPR, 2026
Diffusion-model inversion for reconstruction-free visual anomaly detection.
Shunsuke Sakai
Shunsuke is a Ph.D. student at the University of Fukui, working on artificial intelligence, computer vision, and robust visual representation learning. His research interests include visual anomaly detection, self-supervised learning, diffusion models, dataset distillation, and model reliability for practical recognition systems.
He is a member of Hasegawa Laboratory, where he studies learning methods that improve generalization under limited supervision, distribution shifts, and real-world deployment constraints.
Email / Google Scholar / X / LinkedIn
News
- 2026-10-5: Started a six-month visit to UBC Vancouver as a Visiting International Research Student (VIRS), working with Leonid Sigal and Xiangteng He.
- 2026-09-30: Awarded the JSPS Research Fellowship for Young Scientists (DC2), April 2027 – March 2029.
- 2026-08-17: Our co-first-author paper DSeq-JEPA was accepted to ECCV 2026. Congrats to Xiangteng He!
- 2026-06-05: Presented my first-author paper InvAD at CVPR 2026. Huge thanks to Chunzhi Gu and Xiangteng He!
- 2026-06-01: unCLIP-KD was accepted to Neurocomputing (IF: 6.7). Congrats to Tatsuya Sasaki!
- 2026-06-01: Released Pool-Select-Refine, an allocation-aware approach to generative dataset distillation.
- 2026-04-01: Started my Ph.D. at the University of Fukui under the supervision of Tatsuhito Hasegawa.
- 2026-03-14: Released VID-AD, a logical anomaly detection dataset featuring vision-induced distractions.
- 2026-03-13: My first-author paper LADMIM was accepted to TMLR. Huge thanks to my co-authors!
- 2026-02-15: Our work on few-shot human action anomaly detection was published in Knowledge-Based Systems (IF: 8.0). Congrats to Koichiro Kamide!
- 2025-11-15: Two papers were published at ICONIP 2025. Congrats to Wenmin Li and Tatsuhito Hasegawa!
- 2025-08-02: Our work on analytical softmax temperature setting was accepted to Neural Computing and Applications (IF: 4.5). Congrats to Tatsuhito Hasegawa!
Older news
- 2024-10-14: Released my first-author preprint LADMIM on logical anomaly detection with masked image modeling.
Research
I am broadly interested in computer vision and machine learning methods that remain useful when data are scarce, noisy, shifted, or difficult to annotate.
DSeq-JEPA: Discriminative Sequential Joint-Embedding Predictive Architecture
Xiangteng He, Shunsuke Sakai, Kun Yuan, Nicolas Padoy, Tatsuhito Hasegawa, Leonid Sigal
ECCV, 2026
ECCV poster / arXiv / code
Recasting JEPA target prediction as sequence learning for transferable visual representations.
Logical Anomaly Detection with Masked Image Modeling
Shunsuke Sakai, Tatsuhito Hasegawa, Makoto Koshino
TMLR, 2026
OpenReview / arXiv / code
Masked image modeling in a discrete latent space for logical anomaly detection.
unCLIP-KD: A generative knowledge distillation framework conditioned on CLIP representations
Tatsuya Sasaki, Shunsuke Sakai, Tatsuhito Hasegawa
Neurocomputing, 2026
Generative knowledge distillation from CLIP-conditioned image variants.
Pool-Select-Refine: Allocation-Aware Generative Dataset Distillation with Soft-Label-Guided Latent Refinement
Wenmin Li, Shunsuke Sakai, Zhongkai Zhao, Tatsuhito Hasegawa
arXiv preprint, 2026
Allocation-aware generative dataset distillation with candidate selection and latent refinement.
VID-AD: A Dataset for Image-Level Logical Anomaly Detection under Vision-Induced Distraction
Hiroto Nakata, Yawen Zou, Shunsuke Sakai, Shun Maeda, Chunzhi Gu, Yijin Wei, Shangce Gao, Chao Zhang
arXiv preprint, 2026
Logical anomaly detection under controlled clutter, illumination shift, and blur.
Few-shot human action anomaly detection via a unified contrastive learning framework
Koichiro Kamide, Shunsuke Sakai, Shun Maeda, Chunzhi Gu, Chao Zhang
Knowledge-Based Systems, 2026
Few-shot action anomaly detection with contrastive learning and generative motion augmentation.
Contrastive Learning-Enhanced Trajectory Matching for Small-Scale Dataset Distillation
Wenmin Li, Shunsuke Sakai, Tatsuhito Hasegawa
Neural Information Processing (ICONIP 2025), 2026
Contrastive trajectory matching for informative tiny distilled datasets.
Personalization of Human Activity Recognition with Cross-Conditioned Diffusion Models
Tatsuhito Hasegawa, Shunsuke Sakai
Neural Information Processing (ICONIP 2025), 2026
Cross-conditioned diffusion for subject-adapted activity recognition waveforms.