The moon over Melbourne

Xiaohaoyang Lei

1Monash University2Southeast University–Monash University Joint Graduate School

xlei0009@student.monash.edu

Abstract

I’m an M.Sc. student at Monash University working on generative models for image editing and video generation. Lately I’ve been drawn to action-conditioned video world models, which predict how a scene unfolds given what an agent does.

Index Terms—Image editing, video generation, diffusion models, world models.

Xiaohaoyang Lei with arms outstretched in front of a Monash University billboard
Figure 1. Reverse diffusion from t = 1000 to 0, ending at my portrait. Every visit starts from fresh noise.

1News

  1. May 2026

    BlendEmo, our co-first-author paper on blended emotion recognition, was presented as an oral at IEEE FG 2026. Paper

  2. 2025

    Started the M.Sc. in Data Science at Monash University, and joined the Southeast University–Monash University Joint Graduate School as a research assistant.

  3. 2025

    Teaching assistant for Fundamentals of Artificial Intelligence, Machine Learning, and Natural Language Processing.

  4. 2024

    Graduated from the University of Liverpool with a B.Eng. in Electrical Engineering and Electronics.

Earlier news (3)
  1. Sep 2023

    Our monolithic GaN DC–DC floating buck converter was published in Semiconductor Science and Technology. DOI (opens in a new tab)

  2. 2023

    Presented R2I, my first-author paper on photovoltaic fault diagnosis, as an oral at IEEE-EPEE 2023. Paper

  3. 2022

    Best Poster Award at the International Forum on Wide Bandgap Semiconductors (IFWS).

2Publications

* Equal contribution. Hover or tap a figure to see the task it tackles. Full list on Google Scholar (opens in a new tab).

  1. BlendEmo: Structured set prediction and pair-conditioned ratio modeling for blended emotion recognition

    K. Cheng*, X. Lei*, R. Wang, H. Li, and C. Chen

    IEEE International Conference on Automatic Face and Gesture Recognition (FG), 2026OralCo-first author

    Recognises which two emotions are blended in a face and how much each contributes, using structured set prediction with pair-conditioned ratio decoding.

  2. R2I: A deep fault diagnosis semantic segmentation system based on knowledge distillation in photovoltaic panels

    X. Lei, K. Ma, and J. Zhao

    International Conference on Energy, Power and Electrical Engineering (IEEE-EPEE), 2023OralFirst author

    Distils an RGB U-Net teacher into a lightweight infrared student to segment panel faults, then estimates severity from the predicted defect area.

  3. A 35 V–5 V monolithic integrated GaN-based DC–DC floating buck converter (opens in a new tab)

    Y. Zhu, H. Su, F. Li, A. Li, X. Lei, M. Cui, W.-C. Lai, H. Wen, and W. Liu

    Semiconductor Science and Technology, vol. 38, no. 9, 095001, 2023Journal

    A GaN power IC that steps 35 V down to 5 V on a single chip. I worked on device characterisation, converter measurement and the paper's figures.

    DOI (opens in a new tab)

  4. Review on the FPGA design optimization for ASIC

    X. Lei

    International Conference on Software Engineering and Machine Learning (ICSEML), vol. 4, pp. 524–531, 2023Sole author

    Surveys how FPGA prototyping and optimisation techniques carry over into ASIC design flows.

3Research

Now. Generative models for images and video. I’m interested in edits that change exactly what an instruction asks for and leave everything else alone, and in video generation that stays consistent from one frame to the next.

Next. Action-conditioned video world models, which roll a scene forward frame by frame given an agent’s actions: a step from generating video towards simulating the world it shows. Figure 2 is a toy version.

Before. Blended emotion recognition with structured set prediction [1]BlendEmo: Structured set prediction and pair-conditioned ratio modeling for blended emotion recognitionIEEE FG 2026, RGB-to-infrared distillation for photovoltaic fault segmentation [2]R2I: A deep fault diagnosis semantic segmentation system based on knowledge distillation in photovoltaic panelsIEEE-EPEE 2023, and, as an undergraduate across XJTLU and Liverpool, measuring GaN power ICs [3]A 35 V–5 V monolithic integrated GaN-based DC–DC floating buck converterSemiconductor Science and Technology 2023.

frame=0action=nonesteps=8

Figure 2. A toy action-conditioned world model. Choose an action (or use the arrow keys) and the next frame is denoised from noise, conditioned on it. The “model” is a hand-written renderer, not a trained network, so it never hallucinates.

4Background

Table 1. Education, experience, teaching and honours.
PeriodTitleInstitution or venue
Education
2025 – nowM.Sc. Data ScienceMelbourne, AustraliaMonash University
2022 – 24B.Eng. Electrical Engineering and Electronics2+2 joint programme, Liverpool, UKUniversity of Liverpool
2020 – 22B.Eng. Telecommunications Engineering2+2 joint programme, Suzhou, ChinaXi’an Jiaotong-Liverpool University
Research and industry
2025 – nowResearch AssistantComputer vision, affective computing and multimodal learningSoutheast University–Monash University Joint Graduate School
2024Trainee Engineer InternIoT RF modules, test PCBs and an emergency-broadcast applicationBEHOME
2022 – 23Research AssistantGaN device measurement, data processing and figuresXJTLU Nano-Micro Fabrication Laboratory
Teaching
2025 – nowTeaching assistantFIT5047 Fundamentals of Artificial Intelligence; FIT5201 Machine Learning; FIT5217 Natural Language ProcessingSoutheast University–Monash University Joint Graduate School
Honours
2026Oral presentationIEEE FG
2023Oral presentationIEEE-EPEE
2022Best Poster AwardIFWS

AVisitors

Where readers of this page come from. Counts are page views, not people.

—Total page views
—Countries / regions
—Unknown location
Dotted world map

Loading visitor statistics…

Earlier visitsVisited in the last 24 hoursPoint size reflects page views
Regional statistics

Loading regional statistics…

Hover, focus, or tap a point for local page views. Overlapping dots share a point with separate counts. Totals include repeat visits and visits without a detailed location; locations are approximate.

Statistics hosted on Cloudflare. A new record begins with this migration. Map data: Natural Earth (opens in a new tab).