Projects
Flagship Projects
DataWink supports creative reuse, using a VLM-powered pipeline to extract a template from a vis example and allows dynamic adaptation in low-level features.
WaitGPT translates complex LLM-generated code into a live, interactive diagram, giving data analysts the power to proactively steer the agent’s analysis.
Philosophy
My research is driven by a core belief: tools must fully respect our creative vision. I design systems using a framework I call “intent-first, nuances-on-demand”, which respects our vision at every stage, mapping high-level intent into a manipulatable artifact for iterative refinement over the nuances.
Read more: Problems with existing creative tools
- Classic tools like PhotoShop or Excel are masters of nuance, but disrespect high-level intent. They force us to do all the translation, manually breaking the vision into a thousand technical steps.
- GenAI tools like Midjourney or Tableau's Ask Data seem to respect high-level intent, but they disrespect the evolving, steering intent. They offer poor steerability and no path for refinement, forcing us to accept their version of our own idea.
My projects apply this framework to generative creative tools, such as graphics/animation tools and explorative analysis tools. This domain is the perfect testbed, as it requires balancing two distinct types of creative vision: expressive, narrative intent and logical, analytical intent.
Respecting Expressive, Narrative Intent
I study mostly on the expressive, narrative side of intent. These projects include empirical studies to understand why creator intent gets compromised, and authoring tools designed to respect storytelling goals.
DataSway brings metaphorical visualizations to life, translating a creator’s high-level animation intent into vivid motion that can be tweaked with low-level clips.
Emordle generates word clouds that express nuanced emotion, moving beyond simple categories to allow for subtle and personal refinement with dedicated parameters.
This project explores digital collages creation for storytelling. Driven by story description, the tool supports choosing, cutting out, and arranging photo segments.
An interview study on the creative workflows, challenges, and opportunities of amateur creators in authoring animated VR stories.
GeoCamera is an authoring tool for specifying camera effects in geospatial data videos according to narrative purposes
Early Explorations in Visual Analysis
My early work explores how visualization systems can better assist in data comprehension. This field is personally dear to me, because it uniquely embodies both insight discovery and storytelling, which is a duality that inspired my passion for expressivity.
Through these projects, I learned the power of the classic mantra: “overview first, zoom and filter, details on demand”. This idea of structured, user-driven refinement is not just a tool for rigid systems; it provides the exact principles needed to control and steer modern generative creative tools.
OldVisOnline is a digital library with 13K+ vis images created before the 1950s. It offers a fresh perspective on the history of vis.
AutoCaption is one of the early works in vis to generate natural language-based descriptions for charts, capturing its visual features.
Wordle Duet encodes hierarchical text as a dual resolution wordle to bring the context together with keywords, inspired by hybrid images.
MeasureFlow presents a series of graph measures as an overview of a dynamic network for exploration. It is an extension of Vistorian.
PIMA encodes group features onto a sequence of matrices with pixel technique and supports interactions in a top-down manner.