Hi there! I am a 5th-year CS PhD candidate at Stony Brook University, advised by Prof. Arie Kaufman. Currently, I am a student researcher at ByteDance Seed, working on video generation with Peng Wang, Yichun Shi, and Ceyuan Yang. Previously, I interned in the Video AI group at Adobe Research in Summer 2024, working on video generation with Yang Zhou and Prof. Feng Liu. I was a research intern in the 3D group at Adobe Research in Summer 2023, working with Xin Sun and Hao Tan. Prior to joining SBU, I spent 3 years at Lehigh University for my undergraduate CS education, where I worked on protein volume classification with Prof. Brian Chen.

I am broadly interested in visual generative models, large-scale training, and Reinforcment Learning (RL). In my research, I often take a scalability perspective (e.g. the Bitter Lesson):

• Progressive Autoregressive Video Diffusion Models introduces a new noise schedule and sampling algorithm that seamlessly model long videos with regular context length; otherwise, long videos are computationally infeasible to diffusion modeling directly.
• Carve3D uses RL with feedback from 3D reconstruction to circumvent using real 3D data.
• LRM-Zero introduces large-scale procedural 3D data synthesis to address 3D data scarcity.

News

03/24/2025 Started internship at ByteDance Seed, working on video generation.

12/10/2024 Attended NeurIPS 2024 in Vancouver, BC, Canada. I enjoyed many great social events, hosted by Google, Meta, Luma, World Labs, Genmo, and Playground.

10/10/2024 Progressive Autoregressive Video Diffusion Models is released on arXiv. Proud to build the first video generation model that can autoregressively generate 60-second videos without quality degradation over time.

09/25/2024 LRM-Zero is accepted to NeurIPS 2024.

06/17/2024 Attended CVPR 2024 in Seattle, WA. I presented Carve3D and was glad to see LRM-Zero being featured in two workshop invited talks (3DFM and SyntaGen) by Hao Tan and Nathan Carr.

06/13/2024 LRM-Zero is released on arXiv. First time working on large-scale pre-training and data generation, and it was a blast!

05/28/2024 Started my second internship at Adobe Research, this time working on video generation.

02/26/2024 Carve3D is accepted to CVPR 2024.

07/17/2023 GAIT is accepted to ICCV 2023.

06/19/2023 Started my internship at Adobe Research.

Publications

Progressive Autoregressive Video Diffusion Models
Desai Xie, Zhan Xu, Yicong Hong, Hao Tan, Difan Liu, Feng Liu, Arie Kaufman, Yang Zhou
CVPR 2025 CVEU Workshop
Project Paper Code

LRM-Zero: Training Large Reconstruction Models with Synthesized Data
Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Sören Pirk, Arie Kaufman, Xin Sun, Hao Tan
NeurIPS 2024
Project Paper Code

Carve3D: Improving Multi-view Reconstruction Consistency for Diffusion Models with RL Finetuning
Desai Xie, Jiahao Li, Hao Tan, Xin Sun, Zhixin Shu, Yi Zhou, Sai Bi, Sören Pirk, Arie E. Kaufman
Conference on Computer Vision and Pattern Recognition (CVPR), 2024
Project Paper Code

GAIT: Generating Aesthetic Indoor Tours with Deep Reinforcement Learning
Desai Xie, Ping Hu, Xin Sun, Sören Pirk, Jianming Zhang, Radomír Měch, Arie E. Kaufman
International Conference on Computer Vision (ICCV), 2023
Project Paper Code

Misc

A fun fact about my name is that De (德) and Sai (赛) means demoncracy and science in Chinese (Wikipedia).

I love training my “Catificial” Intelligence/CatGPT🐈 agent, Purrari, using a blend of supervised learning (instruction finetuning) and RL (treats as positive reward). She understands many words in both English and Mandarin and has mastered numerous tricks. Currently, she is advancing her communication skills through talking buttons. For more cute cat pics and videos, please visit her instagram, lovingly maintained by her mom.