Ji Xie

I am a first-year PhD student at CMU LTI, advised by Prof. Louis-Philippe Morency. Previously, I was a research intern at ByteDance Seed and a visiting student at the Berkeley AI Research (BAIR) Lab, UC Berkeley, advised by Prof. XuDong Wang and Prof. Trevor Darrell.

My research focuses on scalable, unified, controllable multimodal models, with a long-term goal of developing world models for embodied intelligence.

Email / Google Scholar / CV / GitHub / Twitter

Industry Experience

Seedream 5.0 series

Core contributor to interactive editing, including any-color control, layer separation, text edit, and spatial grounding for precise local editing.

tech blog / project page

Favorite Works

Foundation Models
	Reconstruction Alignment Improves Unified Multimodal Models Ji Xie, Trevor Darrell, Luke Zettlemoyer, Xudong Wang ICLR 2026 paper / code / model Unlocking the massive zero-shot potential in unified multimodal models through self-supervised learning.
	Seeing Sound, Hearing Sight: Uncovering Modality Bias and Conflict of AI Models in Sound Localization Yanhao Jia, Ji Xie, S. Jivaganesh, Hao Li, Xu Wu, Mengmi Zhang NeurIPS 2025 (Spotlight) paper / code SSHS studies modality conflict in sound localization and shows EchoPin improves robustness under conflicting cues.

Application
	Paint-Anything: Toward Any-Color Controllable Image Generation and Editing Ji Xie, Dewei Zhou, Xinyu Huang, Xun Wang Seed Technical Report 2026 Paint-Anything enables unified, prompt-native 24-bit hex-color control for both image generation and editing in a single model.
	MetaPoint: Unlocking Precise Spatial Control in Agentic Visual Generation Dewei Zhou, Xinyu Huang, Xun Wang†, Ji Xie, Yabo Zhang, Liang Li, Kunchang Li, Zongxin Yang, Yi Yang denotes equal contribution, † denotes project lead ECCV 2026 paper MetaPoint enables precise point-based spatial control for agentic visual generation.
	VideoCoF: Unified Video Editing with Temporal Reasoner Xiangpeng Yang, Ji Xie, Yiyuan Yang, Yue Ma, Yan Huang, Min Xu, Qiang Wu CVPR 2026 (Highlight) paper / code / model / project page VideoCoF uses a see-reason-edit pipeline for mask-free, precise video editing and strong long-video extrapolation.
	In-Context Edit: Enabling Instructional Image Editing with In-Context Generation in Large-Scale Diffusion Transformer Zechuan Zhang, Ji Xie, Yu Lu, Zongxin Yang, Yi Yang NeurIPS 2025 paper / code (2K Stars🌟) / model Image editing is worth a single LoRA! With only 0.1% training data, ICEdit delivers fantastic instructional image editing.
	3DIS: Depth-Driven Decoupled Instance Synthesis for Text-to-Image Generation Dewei Zhou, Ji Xie, Zongxin Yang, Yi Yang *( denotes equal contribution) ICLR 2025 (Spotlight)** paper / code / model 3DIS uses depth-driven decoupled instance synthesis for controllable text-to-image generation.

Experience

	First-year PhD Student, Language Technologies Institute, Carnegie Mellon University Advisor: Louis-Philippe Morency August 2026 ~ Present
	Research Intern, ByteDance Seed Mentor: Xun Wang October 2025 ~ August 2026
	Visiting Student, Berkeley AI Research (BAIR), UC Berkeley March 2025 ~ December 2025

Invited Talks

"Reconstruction Alignment Improves Unified Multimodal Models"
Apple Research · Invited Talk · Hosted by Chen Chen and Yinfei Yang
October 2025

Selected Honors & Awards

SenseTime Scholarship
Top 30 recipients annually in China
June 2025

Zhejiang Provincial Higher Mathematics Competition, First Prize
June 2024

Zhejiang Provincial Collegiate Programming Contest, Gold Medal
April 2024, April 2023

International Collegiate Programming Contest (ICPC), Shenyang Site Gold Medal
October 2022

China Collegiate Programming Contest (CCPC), Guangzhou Site Gold Medal
October 2022

Miscellaneous

I have competitive-programming experience in ACM/ICPC and achieved a rating of 2478 on Codeforces. You can find my old blog here — it contains my competitive-programming notes :)

Website template from Jon Barron