📄 Curriculum Vitae

MIT HAN Lab Intern

Apr. 2025 – Present

Conducting research on generative model acceleration at MIT HAN Lab under the guidance of PhD student Muyang Li.

• Core contributor to nunchaku (2.6K⭐) and ComfyUI-nunchaku (1.8K⭐) Integrated PuLID into nunchaku, enabling extreme acceleration and memory compression for Flux1.dev-based PuLID with minimal quality loss
• Developed corresponding ComfyUI nodes
• Conducting research on quantization training of diffusion models

AIGC, Diffusion, Quantization, Acceleration

Tencent Intern

Nov. 2024 – Sept. 2025

Worked at WXG, researching and implementing the latest image/video face-swapping algorithms.

• First author of Stand-In: A Lightweight and Plug-and-Play Identity Control for Video Generation, achieving superior facial similarity and naturalness with significant efficiency gains. Released code and project page. Supports face swapping, non-human identity-preserving generation, stylized video generation, and pose-driven generation
• Designed new face fusion algorithm for WeChat Channels, surpassing IP-Adapter, InstantID, PuLID; reduced inference time from 20s to 5s
• Performed large-scale fine-tuning of SDXL for high-quality Asian portraits

T2I, T2V, IP2V

IDEA ReadPaper Intern

Aug. 2024 – Nov. 2024

Designed and built academic graph systems and disambiguation algorithms for LLM reasoning and paper recommendation.

• Developed knowledge graph schema for academic resources
• Built entity disambiguation and deduplication systems
• Integrated into production at www.readpaper.com

Graph datasets, RAG

LoRAExpand

Nov. 2023 – Jun. 2024

Built a model capable of generating landscape paintings while placing specific calligraphy in designated regions.

• Trained LoRA with Chinese landscape dataset
• Integrated SVD for text-to-video rendering
• Combined with AnyText for dynamic calligraphy placement
• Deployed UI with Gradio
• Code: GitHub

Diffusion, LoRA, AnyText, Gradio

Deepin AI QAbot

May 2024 – Aug. 2024

Built a lightweight RAG system to answer domain-specific queries over structured documents.

• Designed RAG pipeline with custom retriever + reader + answerer
• Deployed with Gradio UI and minimal hardware requirement

LLM, RAG, Gradio