Shengqu Cai

(Shengqu) Cai 「蔡盛曲」

You may wonder why "Prime", which is such a strange nickname, so here is the story: I have been having a deep voice since middle school, and that is when I earned myself the nickname "Prime" (obviously because of the famous character who's known to have a deep voice too). So yes, this nickname is earned, not something that randomly popped into my head.

I'm a CS PhD student at Stanford University, advised by Prof. Gordon Wetzstein and Prof. Leonidas Guibas, affliated with Computational Imaging Lab and Geometric Computing Lab. I am partly supported by a Stanford School of Engineering Fellowship.

Before Stanford, I was a CS master student at ETH Zürich supervised by Prof. Luc Van Gool. I obtained my Bachelor degree in Computer Science with first honour from King's College London in United Kingdom, where I spent some time working on information theory.

In 2022, I spent a few wonderful months working on diffusion with Eric Chan and Songyou Peng. I started my research career back in 2021 working on NeRFs and GANs with Anton Obukhov. I consider them as mentors coming into research and who I try to learn from.

I am interested in solving tasks that are fundamentally ill-posed via traditional methods, slay the unslayable. I have been working around generative models, primarily on video generation these days, hoping they can one day forward simulate the future, and reverse engineer the past. I like making cool theories, videos, demos and applications.

Email / CV / Google Scholar / Semantic Scholar / Github / Twitter / Linkedin

* This is me prior-COVID. Since then I gained >40 pounds and lost my cool
;(

News

2026-02: BulletTime is accepted to CVPR 2026!
2026-02: Generative Reality is accepted to CVPR 2026 Findings!
2026-01: MoC and Captain Cinema are accepted to ICLR 2026!
2025-09: I wrote a short blog on long-context memory of video world models.
2025-09: FramePack is accepted to NeurIPS 2025 as a spotlight!
2025-06: Joining NVIDIA Research this summer!
2025-03: CL-Splats is accepted to ICCV 2025!
2025-03: We received an NVIDIA Grant for our research!
2025-02: DSD and X-Dyna are accepted to CVPR 2025!
2024-09: CVD is accepted to NeurIPS 2024!
2024-02: Generative Rendering is accepted to CVPR 2024!
2023-09: I joined Stanford University for PhD in Computer Science!
2023-07: DiffDreamer is accepted by ICCV 2023!
2023-05: I graduated from ETH Zürich!
2023-01: I will be working as a research intern at Adobe this summer!
2022-03: Pix2NeRF is accepted by CVPR 2022. First submission first accept!

Publications

* and † indicate equal contribution

	Mode Seeking meets Mean Seeking for Fast Long Video Generation Shengqu Cai, Weili Nie, Chao Liu, Julius Berner, Lvmin Zhang, Nanye Ma, Hansheng Chen, Maneesh Agrawala, Leonidas Guibas, Gordon Wetzstein, Arash Vahdat In arXiv 2026 [Project Page][Paper] Decouple local realism and long-range coherence for fast long-video generation by combining mode-seeking teacher distillation with mean-seeking long-video supervision.
	Mixture of Contexts for Long Video Generation Shengqu Cai, Ceyuan Yang, Lvmin Zhang, Yuwei Guo, Junfei Xiao, Ziyan Yang, Yinghao Xu, Zhenheng Yang, Alan Yuille, Leonidas Guibas, Maneesh Agrawala, Lu Jiang, Gordon Wetzstein In ICLR 2026 [Project Page][Paper] Learnable sparse attention routing enables minute-long context with short-video cost, pruning most token pairs while preserving minute-long context coherence.
	FramePack: Frame Context Packing and Drift Prevention in Next-Frame-Prediction Video Diffusion Models Lvmin Zhang, Shengqu Cai, Muyang Li, Gordon Wetzstein, Maneesh Agrawala In NeurIPS 2025 (Spotlight) [Project Page] [Paper] [Code] Frame context packing for next-frame prediction enables longer contexts within fixed sequence budgets, with drift prevention to reduce error accumulation.
	BulletTime: Decoupled Control of Time and Camera Pose for Video Generation Yiming Wang, Qihang Zhang, Shengqu Cai, Tong Wu†, Jan Ackermann†, Zhengfei Kuang†, Yang Zheng†, Frano Rajič†, Siyu Tang, Gordon Wetzstein In CVPR 2026 [Project Page][Paper] Purely implicit time- and camera-controlled 4D video diffusion that enables decoupled control over world time and camera pose (bullet-time effects).
	Generated Reality: Human-centric World Simulation using Interactive Video Generation with Hand and Camera Control Linxi Xie, Lisong C. Sun, Ashley Neall, Tong Wu, Shengqu Cai, Gordon Wetzstein In CVPR Findings 2026 [Project Page] [Paper] Interactive human-centric video world simulation conditioned on tracked head pose and joint-level hand poses.
	Diffusion Self-Distillation for Zero-Shot Customized Image Generation Shengqu Cai, Eric Ryan Chan, Yunzhi Zhang, Leonidas Guibas, Jiajun Wu, Gordon Wetzstein In CVPR 2025 [Project Page][Paper][Code][Demo] Training-free customized image generation model that scales to any instance and any context.
	Captain Cinema: Towards Short Movie Generation Junfei Xiao, Ceyuan Yang, Lvmin Zhang, Shengqu Cai, Yang Zhao, Yuwei Guo, Gordon Wetzstein, Maneesh Agrawala, Alan Yuille, Lu Jiang In ICLR 2026 [Project Page] [Paper] Top-down keyframe planning with bottom-up long-context video synthesis, using interleaved training to adapt MM-DiT for stable and efficient multi-scene generation.
	CL-Splats: Continual Learning of Gaussian Splatting with Local Optimization Jan Ackermann, Jonas Kulhanek, Shengqu Cai, Haofei Xu, Marc Pollefeys, Gordon Wetzstein, Leonidas Guibas, Songyou Peng In ICCV 2025 [Project Page][Paper][Code] Efficiently updates Gaussian splatting-based 3D scene reconstructions from incremental images via change detection and local optimization.
	ByteMorph: Benchmarking Instruction-Guided Image Editing with Non-Rigid Motions Di Chang, Mingdeng Cao, Yichun Shi, Bo Liu, Shengqu Cai, Shijie Zhou, Weilin Huang, Gordon Wetzstein, Mohammad Soleymani, Peng Wang In arXiv 2025 [Project Page] [Paper] [Code] Large-scale benchmark and baseline for instruction-guided image editing with complex non-rigid motions such as viewpoint changes, articulations, and deformations.
	ReStyle3D: Scene-level Appearance Transfer with Semantic Correspondences Liyuan Zhu, Shengqu Cai, Shengyu Huang, Gordon Wetzstein, Naji Khosravan, Iro Armeni In SIGGRAPH 2025 [Project Page][Paper][Code] Scene-level appearance transfer from a single style image to multi-view real-world scenes with semantic correspondences.
	X-Dyna: Expressive Dynamic Human Image Animation Di Chang, Hongyi Xu, You Xie, Yipeng Gao, Zhengfei Kuang, Shengqu Cai, Chenxu Zhang, Guoxian Song, Chao Wang, Yichun Shi, Zeyuan Chen, Shijie Zhou, Linjie Luo, Gordon Wetzstein, Mohammad Soleymani In CVPR 2025 (Highlight) [Project Page][Paper][Code] Human image animation using facial expressions and body movements derived from a driving video.
	Collaborative Video Diffusion: Consistent Multi-video Generation with Camera Control Zhengfei Kuang, Shengqu Cai, Hao He, Yinghao Xu, Hongsheng Li, Leonidas Guibas, Gordon Wetzstein In NeurIPS 2024 [Project Page][Paper][Code] Multi-view/multi-trajectory generation of videos sharing the same underlying content and dynamics.
	Robust Symmetry Detection via Riemannian Langevin Dynamics Jihyeon Je, Jiayi Liu, Guandao Yang, Boyang Deng, Shengqu Cai, Gordon Wetzstein, Or Litany, Leonidas Guibas In SIGGRAPH Asia 2024 [Project Page][Paper] Render low fidelity animated mesh directly into animation using pre-trained 2D diffusion models, without the need of any further training/distillation.
	Generative Rendering: Controllable 4D-Guided Video Generation with 2D Diffusion Models Shengqu Cai, Duygu Ceylan, Matheus Gadelha, Chun-Hao Paul Huang, Tuanfeng Y. Wang, Gordon Wetzstein In CVPR 2024 [Project Page][Paper] Render low fidelity animated mesh directly into animation using pre-trained 2D diffusion models, without the need of any further training/distillation.
	DiffDreamer: Towards Consistent Unsupervised Single-view Scene Extrapolation with Conditional Diffusion Models Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, Gordon Wetzstein In ICCV 2023 [Project Page][Paper][Code] A diffusion-model based unsupervised framework capable of synthesizing novel views depicting a long camera trajectory flying into an input image.
	Pix2NeRF: Unsupervised Conditional π-GAN for Single Image to Neural Radiance Fields Translation Shengqu Cai, Anton Obukhov, Dengxin Dai, Luc Van Gool In CVPR 2022 [Paper][Code] 3D-free unsupervised Single view NeRF-based novel view synthesis via conditional NeRF-GAN training and inversion.

Misc

Conference Review: CVPR, ICCV, ECCV, NeurIPS, ICLR, ICML, Eurographics, SIGGRAPH

Journal Review: IJCV, Computing Surveys