2022.09 — 2025.06
Tsinghua University
Tsinghua Shenzhen International Graduate School
Haojun Zhang (Elgin)张昊钧
Research interests: Multimodal Learning, Omni Models, and Generative AI.
Currently open to potential work opportunities and academic collaboration. Please feel free to contact me by email if you are interested!
Education
Education
2017.09 — 2021.06
Nanjing University
School of Physics
Experience
Work Experience
2025.06 — Present
XPENG Inc.
Developing foundation-model capabilities for intelligent cockpits and on-device deployment, spanning VLM-based GUI agents, lightweight audio encoders for omni models, and low-latency streaming ASR. Building iterative pipelines across data, training, evaluation, and efficient inference.
Selected Work
Projects
2025.01 — 2025.04
Identity-Preserving Full-Body Style Transfer
Designed a commercial AIGC pipeline from character style transfer to identity injection using FLUX.1 Redux and Fill, SegFormer, block-wise inpainting, style LoRA, and ControlNet. Investigated layer-wise content and style injection within the FLUX.1 MMDiT architecture.
2024.08 — 2025.02
AIGC Advertising Asset Generation
Built an advertising-content understanding and prompt-expansion pipeline with VLMs and LLMs. Aligned generation models with business concepts using LoRA and DDPO, then combined FLUX.1 Redux and ControlNet for identity-consistent generation and style transfer across more than 20 banner categories.
Research
Publication
X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation
A progressive audio-encoder compression framework that uses behavioral probing, representation alignment, and cross-scale distillation to reduce inference cost while preserving—and in selected configurations improving—speech recognition performance.