Qiucheng Wu

I am a final-year Ph.D. student in Computer Science at the University of California, Santa Barbara, under the supervision of Prof. Shiyu Chang. I received my bachelor’s and master’s degrees from the University of Michigan, College of Engineering, and a second bachelor’s degree at Shanghai Jiao Tong University.

My research focuses on generative models at the intersection of vision and language. Specifically, I am interested in:

• The imbalanced capability of multimodal large language models (MLLMs) in understanding visual and textual information, and methods to enhance their performance on visual tasks.
• Text control in visual generative models (e.g., diffusion models, GANs) for images and videos.

Some of my projects and publications are shown below.

Email: qiucheng@ucsb.edu /  CV

* indicates authors with equal contributions.

To diagnose the imbalance between visual and textual understanding in MLLMs, we introduce VSP, focusing on evaluating their spatial planning ability and further analyzing weaknesses through fine-grained perception and reasoning tasks.

We propose a new text-to-image algorithm with explicit control over cross-attention in diffusion models from spatial and temporal views. This alleviates inconsistencies between images and text and helps to fix errors like missing objects, mismatched attributes, and mislocated objects.

Based on a fixed stable diffusion model, we disentangle target attributes from a single training image. The learned parameters can then be applied to unseen images and achieve the same edits. This finding leads to a lightweight image editing framework with only 50 learnable parameters.

We propose Ada-Deblur, a super-network that can be applied to a "broad spectrum" of various blur levels with no retraining on novel blurs.

Meta reinforcement learning requires substantial amounts of data. To improve the sample efficiency and performance of metal-RL algorithms on sparse-reward tasks, we introduce a novel objective function to learn an action translator among training tasks.

We propose DataSifter-Text to protect privacy of sensitive textual dataset. The DataSifter-Text obfuscates identifiable information in the dataset. It effectively balances between privacy and utility.

We apply the compressive big data analytics (CBDA) to analyze salient features and key biomarkers of a high-dimensional clinical dataset.

Most Likely Health Impact in 4th Annual Symposium Poster Competition in Michigan Institute for Data Science, University of Michigan.

We propose DataSifter to obfuscate sensitive clinical datasets while preserving utility for downstream tasks.

Most Interesting Methodological Advances in 4th Annual Symposium Poster Competition in Michigan Institute for Data Science, University of Michigan.

• Trained efficient multimodal LLM agents for Photoshop/Lightroom workflows.
• Developed text-based controllable image/video generation with diffusion models.

Outcomes: ICCV 2025, ICCV 2023, CVPR 2023

• Explored controllable image generation with GANs for creative editing tools.
• Image restoration through adaptive deblurring methods.

Outcomes: CPAL 2024, TIP 2023

• Developed personalized ML models to improve few-shot food classification accuracy in a nutrition app.

• CS165B: Machine Learning (TA, Fall 2021, Winter 2024, UCSB)
• EECS 492: Intro to Artificial Intelligence (TA, Fall 2020, Winter 2021, UMich)
• Vv156: Applied Honors Calculus (TA, Fall 2016, UM-SJTU Joint Institute)

In this capstone project, we designed a framework to boost web communications between terminal devices and servers by introducing edge servers. We use Kubernetes to organize edge servers as different containers. The edge servers are responsible for computing and sending information to terminal devices, leading to more efficient performance.

This website is built using the source code from Jon Barron's public academic website.