About Me

Started systematically learning skills at the intersection of game graphics and AI since 2022 (Rendering / AI / VFX / Toolchain)
Published work on a controllable video camera preview generation system at ECCV 2026 as first student author
Good English reading and writing skills (IELTS 7.0), able to read cutting-edge English literature and technical docs without barriers
Solid foundations in mathematical statistics and data science, with experience in model training and dataset construction & processing
A certain knowledge reserve in computer graphics, familiar with basic data structures and algorithm principles
Tech otaku with strong self-drive and learning initiative, able to quickly absorb new knowledge and apply it in practice. Love researching the implementation principles of in-game special effects
Familiar with common toon rendering techniques, having attempted to land and implement effects in both Unity and UE5
Hobbies: ACGN enthusiast, Kigurumi, photography, fingerstyle guitar(listen only, don’t play) / electric guitar, and everything related to computer graphics

Skills

Cg/HLSL

Python

C++

UE5

Unity

Blender

Houdini

Work Experience

NetEase Thunder Fire · Fuxi AI Lab 2026.05 — Present

Decision Intelligence Group · AI Algorithm Intern

Participated in Naraka: Bladepoint AI Bot R&D, responsible for upper-level script mining — extracting reusable tactical patterns from game recordings
Explored player intent modeling by abstracting Bot intent labels from behavior sequences and state-event streams to support script decision-making
Built a mining pipeline based on temporal knowledge graphs and clustering; Agent queries the graph to generate scripts for match deployment

Tencent IEG · Interactive Entertainment Group 2025.11 — 2026.05

Domestic Publishing · Eco Development · TA Intern

Independently led UE5 AI smart lighting system R&D, built the in-engine lighting data collection and model training pipeline from scratch, preliminarily validating the model's ability to extract explicit in-engine lighting parameters from GBuffer inputs guided by reference images; also participated in the Agentic system for locating and placing logical light sources in complete environments
Responsible for Arc Raider mobile VFX asset cleanup and specification drafting, and participated in optimizing selected effects
Co-developed a new FPS scene lighting pipeline ensuring correct indoor/outdoor exposure; built a physical light fixture plugin with DataAsset config interface, unifying atmosphere, post-processing, TOD and lighting config workflow

Tengku Network Technology Co., Ltd. 2025.08 — 2025.10

TATD Group · TA Intern

Participated in Project One Up, assisted with external asset onboarding under project specifications
Developed and maintained in-engine texture channel merging/mapping tools and scene material override scanning tools

Education

Shanghai University

Digital Media Creative Engineering · M.Sc.

2024 — 2027

Shanghai University of International Business and Economics

Data Science & Big Data Technology · B.Sc.

2020 — 2024

Portfolio

Controllable Video Camera Preview Generation (ECCV 2026)

Built a complete virtual camera data synthesis pipeline on Unity, including timeline baking tools for three common camera data types, and developed a node-based storyboard video generation system

Published at ECCV 2026 (one of the top three CV conferences) as first author. Research on combining in-engine camera preview with video generation models for controllable workflows. Responsible for the main features of the project: Unity LLM interaction for script & shooting plan generation (RAG/Schema structured), node-based GUI editor, 3D camera trajectory dataset construction, autoregressive & Diffusion-based trajectory generation model training, DPO-based trajectory post-training

UE5 AI Smart Lighting System

Agentic Pipeline

HDBSCAN semantic clustering for preliminary grouping; after reasoning by the clustering Agent, Grounding DINO + Line Trace achieves precise localization of logical light sources. Team Agent orchestration avoids context bloat

Data-Driven Pipeline

Independently led during Tencent IEG internship. SIGGRAPH 2027 Work In Progress

Using image generation models such as LumiNet or GPT Img 2 to transfer the lighting style of a reference image onto the Unlit render of an un-lit scene, then combined with GBuffer geometry information, fed into the subsequent lighting parameter extraction model to reconstruct explicit in-engine light parameters. Built the in-engine lighting data collection and model training pipeline from scratch for AI-assisted lighting

The image-to-lighting-parameter extraction model requires training, so we built a lighting data collection pipeline from scratch, including but not limited to GBuffer and various light Component parameters, achieving Scaling Up via Jitter & OLAT data augmentation
The model is designed as a DETR-like Extractor with a Coarse2Fine two-stage training strategy, using Hungarian matching to compute loss for predicting variable-length unordered sets. Data collection is still ongoing, but the model’s capability has been preliminarily validated

FPS Gameplay

Implemented basic FPS gameplay logic from scratch: damage calculation, reloading, hit / recoil feedback
Basic animation logic, animation blending with Aim Offset, and distinct designs for 5 weapons
Complete health system, Niagara Blood Decal (environmental blood splatter), UMG screen blood stains / hit-direction indicator
RPC network replication; first-person arms & third-person body developed simultaneously, with multicast animation / VFX / audio
Red Dead Redemption 2 - Dead Eye:
Skill activation/deactivation post-process animation, Time Dilation
Maintains an FHitResult TQueue to achieve stable client-side marking UI and execution logic; server & client timer handles ensure correct animation execution

UE5 Scene Lighting

TOD System

Unifies physical atmosphere, procedural sky sphere, and PP exposure control; provides a standardized management workflow

Physical Light Fixture Plugin

Scene lighting pipeline for FPS projects ensuring correct indoor/outdoor exposure; the physical light fixture plugin provides a DataAsset config interface, unifying atmosphere, post-processing, TOD and lighting config workflow

Physics-Field Foliage Interaction

Uses a single-layer RT to simulate a mass-spring field; material samples the simulation result for WPO, achieving more natural bending and rebound
Applies exponential decay along the foliage stem, greatly improving the WPO tearing issue

VFX Works

Black Saber Excalibur

Procedural + texture black-white flash, Modulate fake light, Houdini ground crack

Ice / Crystal

Crystal material, spline-based custom particle ribbon, posterization, shockwave, magic circle

Tutorial Works

Fire Meteor & Hell Attack

Parallax crater, post blur, combo logic, UV shearing…

Thunder & Lightning Bolt

Ribbon, line field, asynchronous damage, Event tail…

Ice Spike & Annihilate

Parallax ice surface, SSS ice spike, complex enemy-searching logic, Spawn Group…

Toon Rendering

UE5.7 Source-Modified Custom Toon Rendering

Added a custom Shading Model with Ramp Atlas shadow transition, ILM AO / Specular
Layered hair Matcap, Rim Light and other common toon rendering features
Added an R16 G16 B16 A16 Toon Buffer with custom MRT Payload to pass data to the Deferred Lighting Pass for subsequent lighting calculation
Spherical parallax eye rendering, with AO-corrected facial SDF shadow (Work In Progress)

Bakes tangent-space smoothed normals and curvature into the model’s UV2 & UV3; Bake tool integrated into the engine right-click menu

Custom-modified global illumination replacing spherical harmonics; flattens normals by modifying Lumen to simulate low-frequency GI, reducing the plastic look in dark areas

Unity URP - Genshin Impact Character Shading

Ramp shadow transition, screen-space equidistant rim light, backface normal outline, SDF face shadow…

Others

Shield Hit Effect

WPO, blueprint hit-response logic

Learning-Based Intelligent LOD Generation for VFX

Initiated by Tencent IEG engine graphics remote project. Explored deep learning and reinforcement learning concepts, learned about VFX performance criteria, attempted theoretical modeling, and learned UE5 plugin development from scratch. Successfully implemented batch Niagara LOD module addition as shown (Slate UI, etc.)

C++ Software Ray Tracing Renderer

Following the GAMES101 course, independently implemented BVH and Monte-Carlo microsurface, then completed the whole renderer via Peter Shirley’s Ray Tracing series, adding Tone Mapping and multi-threading on top