Kunjun Li

I am a first-year PhD student at UC Berkeley, interested in building efficient and scalable AI systems.

Previously, I graduated with honors from the National University of Singapore and spent a term at the University of Washington as an exchange student.

Email LinkedIn GitHub Scholar
Portrait of Kunjun Li

Selected Research

NeurIPS 2025 ScaleKV method overview

Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache

Kunjun Li, Zigeng Chen, Cheng-Yen Yang, Jenq-Neng Hwang

NeurIPS 2025

Scale-aware cache management reduces KV memory by 90% without quality loss, making visual autoregressive generation faster and practical at resolutions up to 4K.

Paper Code
Abstract
Visual Autoregressive modeling offers efficient, scalable next-scale prediction, but its coarse-to-fine process produces a rapidly growing KV cache. ScaleKV identifies scale-specific drafters and refiners from their distinct attention patterns, then gives each group differentiated cache capacity. On the Infinity model family, it preserves pixel-level fidelity while reducing KV-cache memory to 10% of the original.
CVPR 2025 TinyFusion framework overview

TinyFusion: Diffusion Transformers Learned Shallow

Gongfan Fang*, Kunjun Li*, Xinyin Ma, Xinchao Wang (co-first author)

CVPR 2025 Highlight (Top 3%)

An end-to-end learnable depth-pruning framework that halves Diffusion Transformer depth and parameters, delivering 2× faster inference at only 7% of the original training cost.

Paper Code
Abstract
TinyFusion learns which Diffusion Transformer layers to remove by optimizing a pruned model for recoverability after fine-tuning. Differentiable sampling and a co-optimized parameter simulate future fine-tuning during pruning, outperforming importance- and error-based alternatives across DiTs, MARs, and SiTs.
IPSN 2024 PixelGen project banner

PixelGen: Rethinking Embedded Camera Systems for Mixed-Reality

Kunjun Li, Manoj Gulati, Dhairya Shah, Steven Waskito, Shantanu Chakrabarty, Ambuj Varshney

ACM/IEEE IPSN 2024 Best Demo Runner-Up

A low-power camera system that combines monochrome images with sensor signals and diffusion models to visualize information beyond the visible spectrum.

Paper Code
Abstract
PixelGen rethinks mixed-reality camera systems by combining energy-efficient monochrome imagery and low-bandwidth sensors with diffusion-based image generation. The result is a richer representation of the environment that can reveal otherwise invisible physical phenomena.

Education

UC Berkeley seal

University of California, Berkeley

Ph.D. in Computer Science

Berkeley, California
National University of Singapore crest

National University of Singapore

B.S. in Computer Engineering

Singapore
University of Washington logo

University of Washington

Exchange Student

Seattle, Washington

Beyond Research

I feel most at home outdoors: hiking, exploring, and photographing the small moments along the way. I also follow basketball and tennis, and I am always happy to talk about research, sports, or the next trail to try.

Let’s connect