Cheng-Yu (Benjamin) Chiang

I recently graduated from my Master Program at at Carnegie Mellon University studying Electrical and Computer Engineering. I recieved my B.S. in Electrical Engineering from University of California, San Diego in 2023.

I worked on leveraging large scale data for artificial intelligence. During my internships, I worked closely with research and product teams to build scalable data pipelines, resolve data bottlenecks, train machine learning models, and present insights through interactive, user-friendly visualizations.

My passion lies in the intersection of artificial intelligence and electrical engineering, focusing on developing intelligent, scalable, and deployable solutions that utilizes both software and hardware for real-world impact.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

Research

Map It Anywhere (MIA): Empowering Bird's Eye View Mapping using Large-scale Public Data

Map It Anywhere (MIA) is a data engine that leverages Mapillary and OpenStreetMap to create a 1.2 million-pair dataset for Bird’s Eye View (BEV) map prediction, enabling diverse and scalable training data. Models trained using MIA’s dataset achieved 35% better zero-shot performance over existing baselines, demonstrating its potential to improve autonomous navigation.

Projects

GenStreet: Augmenting Street View Generation with Geo-Referenced Data

This project generates realistic First-Person View (FPV) street-view images from segmentation masks and natural language inputs by fine-tuning ControlNet, leveraging Birds Eye View (BEV) maps and zero-shot learning with Llama 3.2 for structural and contextual accuracy. The approach improves realism, structural alignment, and feature accuracy, with applications in robotic simulations, urban planning, and interior design.

Reward Multiverse: A Comprehensive Framework for Diverse Reward Models in Image Generation

Customizing text-to-image diffusion with reinforcement learning and self-supervised reward models to align outputs with specific visual attributes like snow, pixelation, and day-night cycle. The framework introduces sliding-scale modification strength for fine control and unlocks new possibilities for image synthesis and editing. Applications include image editing, style transfer, and data augmentation.

Design and source code from Jon Barron's website