Hi there, I鈥檓 Xinrui (Ryan) Jiang, a master鈥檚 student in EE at Stanford University. Prior to that, I received my B.Eng from Fudan University. During my undergraduate study, I was fortunate to be advised by Professor Zhengzhong Tu at Texas A&M University and Professor Berkin Bilgic at Martinos Center for Biomedical Imaging/Harvard Medical School.

My current interests include AI agents and generative modeling. My earlier research focused on low-level vision, including image super-resolution and quantitative MRI reconstruction. I enjoy starting from real-world needs and observations, framing concrete technical problems, and developing practical methods and systems to address them.

馃摑 Publications

arXiv 2026
sym

Game Sound-Effect Completion with Event-Level Transformation Hints

Xinrui Jiang, Heng Yu

Paper 路 Code

  • We complete a new game skin鈥檚 ability sound effects from the base-skin audio, already-finished target assets, and a design description. A pipeline aligns 5.6K paired ability sequences from League of Legends, and a fine-tuned Stable Audio 3 model uses a signed soft mask to let designers set how much each event should change.
NeurIPS 2025
sym

4KAgent: Agentic Any Image to 4K Super-Resolution

Yushen Zuo, Qi Zheng, Mingyang Wu, Xinrui Jiang, Renjie Li, Jian Wang, Yide Zhang, Gengchen Mai, Lihong V. Wang, James Zou, Xiaoyu Wang, Ming-Hsuan Yang, Zhengzhong Tu

Paper 路 Code 路 Project Page

  • We present 4KAgent, an agentic image super-resolution generalist designed to universally upscale any image to 4K resolution, regardless of input type, degradation level, or domain.
ISMRM 2024
sym

NLCG-Net: A Model-Based Zero-Shot Learning Framework for Undersampled Quantitative MRI Reconstruction

Xinrui Jiang, Yohan Jun, Jaejin Cho, Mengze Gao, Xingwang Yong, Berkin Bilgic

Paper 路 Code 路 Presentation

  • We purposed NLCG-Net, a model-based and data-driven framework achieved via self-supervised learning, which incorporates nonlinear conjugate gradient optimization and Neural Network Regularization in a iterative manner and achieves zero-shot quantitative MRI reconstruction.

馃捇 Internships

  • 2026.06 - 2026.09, Machine Learning Engineer at DoorDash. Agentic editing orchestrator for multi-defect video repair
  • 2025.03 - 2025.08, Machine Learning Engineer at Microsoft. Failure-driven, self-refining LLM-as-a-Judge for large-scale generative media evaluation
  • 2024.01 - 2024.04, Machine Learning Engineer at TikTok/Bytedance. Multimodal content understanding and unsupervised mining for short-form video

馃帠 Selected Awards

  • 2023.09 Suiwei Scholarship (top 1%)
  • 2022.09 National Scholarship (top 1%)

馃専 Fun

In my spare time, I enjoy running (about 3.2 miles each time), building LEGO sets, and hanging out with my American Shorthair cat, Dafu.

馃搶 Projects

Harbor Trajectory Viewer

Harbor Trajectory Viewer

Code

  • A local web app for browsing, auditing and annotating Harbor tasks and ATIF agent trajectories. It replays each run as a chat with a file tree that follows the current step, compares two runs side by side from their first divergence point, and supports per-run and per-step annotation with JSONL export.
Stanford CS 448Z
Cat Chaos scene

Physics-Based Animation and Sound

Code 路 Kitchen Demo 路 Cat Chaos Demo

  • Three audiovisual simulation projects for Stanford CS 448Z. Blender rigid-body scenes export collision and animation events that drive physically based sound synthesis, covering modal impact sound, acceleration noise, bubble acoustics, glass fracture, scraping, and fire.
SumForU

SUMFORU: An LLM-Based Review Summarization Framework for Personalized Purchase Decision Support

Report 路 Code

  • We developed SumForU, a persona-steerable review summarization system built on Thinking Machine鈥檚 Tinker platform. This system fine-tunes LLMs on Amazon 2023 reviews and optimizes for user-aligned summaries via preference-based reinforcement learning.