Weichen Fan

About

Weichen Fan

I am a Second-year PhD student at MMLab@NTU, Nanyang Technological University, under the supervision of Prof. Ziwei Liu. I earned my Bachelor’s degree from the University of Electronic Science and Technology of China (UESTC) and my Master’s degree from the National University of Singapore (NUS). My research centers on video generation, diffusion models, and robotics.

Bachelor’s — UESTCMaster’s — NUSPhD — NTU (in progress)

weichen002 [at] e.ntu.edu.sg GitHub Google Scholar

What Kind of PhD Are You? Ten questions. Four attributes. Twenty-four answers.

Ordered newest first. Every title below resolves to a source.

Order publications Order

Refereed peer-reviewed venue Dashed preprint or technical report

  1. Show the Signal, Hide the Noise: Spectral Forcing for Pixel-Space Diffusion

    arXiv 2026
    Teaser figure for Spectral Forcing.

    arXiv Code

    Pixel space, with nowhere for the noise to hide.

  2. SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture

    Tech report 2026
    Teaser figure for SenseNova-U1.

    arXiv Code

    Understanding and generation under one architecture, which is the hard version.

    Rhetoric objects

    RhetoricTechnical report. Core contributor credit, not first author.

  3. The Prism Hypothesis: Harmonizing Semantic and Pixel Representations via Unified Autoencoding

    ECCV 2026
    Teaser figure for The Prism Hypothesis.

    arXiv Code

    Semantics and pixels held in one autoencoder, stated as a hypothesis rather than a trick.

  4. CFG-Zero*: Classifier-Free Guidance with Zero Prediction

    arXiv 2025
    Teaser figure for CFG-Zero*.

    Project Page arXiv Code

    Guidance reconsidered at the zeroth step. The asterisk is load-bearing.

  5. Vchitect-2.0: Parallel Transformer for Scaling Up Video Diffusion Models

    arXiv 2025
    Teaser figure for Vchitect-2.0.

    Project Page arXiv Dataset Code

    A parallel transformer, scaled up, and the dataset shipped with it.

    Rhetoric objects

    RhetoricPreprint. Nineteen authors; co-first among them.

  6. RepVideo: Rethinking Cross-Layer Representation for Video Generation

    arXiv 2025
    Teaser figure for RepVideo.

    Project Page arXiv Code

    What happens between the layers turned out to be the thing worth rethinking.

  7. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers

    Tech report 2024
    Teaser figure for Lumina-T2X.

    arXiv Code

    Any modality, any resolution, any duration — an ambitious sentence to defend.

    Rhetoric objects

    RhetoricTechnical report, not peer reviewed. Contributor credit.

  8. Link-Context Learning For Multimodal LLMs

    CVPR 2024
    Teaser figure for Link-Context Learning.

    Paper Link Code

    The model is given the link between examples, not only the examples.

  9. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models

    Tech report 2023
    Teaser figure for Sphinx.

    arXiv Code

    Weights, tasks and visual embeddings, mixed jointly rather than stacked.

    Rhetoric objects

    RhetoricTechnical report, not peer reviewed. Contributor credit.

  10. Hierarchy Flow For High-Fidelity Image-to-Image Translation

    arXiv 2022
    Teaser figure for Hierarchy Flow.

    arXiv Code

    Translation that keeps its fidelity by keeping its hierarchy.

  11. SAVAnet: Surgical Action-Driven Visual Attention Network for Autonomous Endoscope Control

    IEEE T-ASE 2022
    Teaser figure for SAVAnet.

    Paper Link

    An endoscope that looks where the surgery is going. The robotics start here.

* denotes equal contribution.

  1. Nanyang Technological University

    Nanyang Technological University

    May 2024 — Nov 2024

    Research Manager

    Singapore

  2. HyperGAI

    HyperGAI

    Oct 2023 — May 2024

    Research Scientist

    Singapore

  3. SenseTime

    SenseTime

    Jun 2022 — Oct 2023

    Computer Vision Researcher

    ShangHai

  4. SenseTime

    SenseTime

    May 2021 — Jun 2022

    Research Intern

    Singapore

The other ledger. It is longer than the first one and it was more fun to write.

3,288.1 hours across 8 titles. 343 of 917 achievements, counted across 7 of the 8.

  1. Tom Clancy’s Rainbow Six Siege

    All 48

    1,558.4 hours48 of 48 achievements

    Steam

    Fifteen hundred hours of listening for a footstep through a ceiling. You do not hear the room any more. You hear what is about to come through it.

    Rhetoric objects

    RhetoricSixty-five days. Not of playing across the years — sixty-five entire days, end to end, in one room.

  2. Age of Empires IV: Anniversary Edition

    489.5 hours76 of 175 achievements

    Steam

    Dark, Feudal, Castle, Imperial. Four ages and a thousand years, compressed into something you can lose an evening to, and you keep going back to lose another one.

  3. Dead by Daylight

    437.5 hours130 of 303 achievements

    Steam

    Something is walking toward you and the sound is getting louder. You have heard it four hundred hours’ worth of times. It still works.

  4. Disco Elysium: The Final Cut

    The source

    274.0 hours

    Steam

    Two hundred and seventy-four hours in a city you could not fix, losing an argument to yourself in twenty-four voices. You are reading one of them now. This entire room is borrowed from that game, and you knew that before you got here.

  5. Total War: THREE KINGDOMS

    250.4 hours37 of 99 achievements

    Steam

    You knew how this one ended before you installed it. Everyone does. Two hundred and fifty hours spent checking whether it would end differently for you.

  6. Mount & Blade II: Bannerlord

    103.4 hours8 of 50 achievements

    Steam

    Eight achievements out of fifty. You were not collecting anything. You were on a horse, going somewhere, and you did not stop to be measured.

    Rhetoric objects

    RhetoricYou did not finish going somewhere. You stopped, and then a PhD happened to you.

  7. Baldur’s Gate 3

    91.0 hours29 of 54 achievements

    Steam

    Ninety-one hours, and better than half the achievements. Not a fast reader. A thorough one — you stopped and talked to everybody.

  8. Crusader Kings III

    83.9 hours15 of 188 achievements

    Steam

    Fifteen of a hundred and eighty-eight. This game keeps score of things nobody sets out to do. The dynasty was the point; the ledger was never going to be filled.

Three records. They will play right here, and they do not load until you ask them to.

  1. Pure

    1990

    The Lightning SeedsCloudcuckooland

    Open in Spotify

    Nineteen ninety, and it still sounds like the first warm afternoon of a year. Nothing in it is trying to hurt you. Do you know how rare that is.

  2. Nightingale

    1997

    YanniTribute

    Open in Spotify

    Not one word in the whole of it, and it still gets the entire story out. Somewhere in there an instrument is pretending to be a bird, and you are letting it.

  3. About You

    2022

    The 1975Being Funny In A Foreign Language

    Open in Spotify

    It wants to be enormous and it is not embarrassed about wanting that. The temperature in the room drops a degree. Play it loud or do not play it.

Email
weichen002 [at] e.ntu.edu.sg
GitHub
github.com/WeichenFan
Scholar
Google Scholar

There is a way through

Arrow keys, or drag.