Scholar
Daquan Zhou
Google Scholar ID: DdCAbWwAAAAJ
Bytedance, US
Citations & Impact
Citations
14,360
H-index
36
i10-index
46
Publications
20
Co-authors
5
list available
Contact
Publications
Resume (English only)
Academic Achievements
- HunyuanVideo: A Systematic Framework For Large Video Generative Models
- StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation
- Magic-Me: Identity-Specific Video Customized Diffusion
- MagicVideo & MagicVideo-V2: Efficient High-Aesthetic Video Generation with Latent Diffusion
- DiffFit: Unlocking transferability of large diffusion models via parameter-efficient fine-tuning (CVPR 2023)
- Expanding small-scale datasets with guided imagination (NeurIPS 2023, Corresponding Author & Project Lead)
- Diffusion probabilistic model made slim (CVPR 2022)
- Dataset Quantization (ICCV 2023)
- Scaling & Shifting Your Features (NeurIPS 2022 Spotlight, Equal First Author)
Research Experience
- Led model pre-training and diffusion algorithm design for HunyuanVideo
- Key contributor to video generation projects including StoryDiffusion, Magic-Me, MagicVideo series
- Pioneered research on parameter-efficient fine-tuning (e.g., DiffFit), small-data expansion, and diffusion model slimming
- Developed PLLaVA: a parameter-free extension of LLaVA from images to videos for dense video captioning
- Proposed Dataset Quantization pipeline achieving 5×–10× training speedup (ICCV 2023)
Background
- Currently a Tenure-track Assistant Professor at Peking University
- Focused on minimizing energy and memory consumption for training and deploying powerful AI algorithms
- Applications include Robotics, AIGC, and Vision-Language-Action (VLA) systems
- Research interests: explainable video representation design and efficient long video generation (both training and inference)
- Strong interest in hardware-algorithm co-design, especially DNN architecture and memory co-optimization
- Ongoing work on model and dataset efficiency for discriminative, generative, and multimodal models
