Imagine a rabbit wearing a camo poncho, or a cow with a chain necklace. Easy, right? As humans, we have no problem using our creativity to visualize things that do not exist because of our inherent understanding of the object. The same hasn’t been so easy for a computer to do. So far, a lot has been accomplished to alter the texture and appearance of an object in its entirety, such as changing the color of the cow or morphing the cow into a pig. However, little has been done to address the generation of detailed stylistic edits on a highly specific region within the object, confining the location of the edit using nothing more than textual prompts. To address this problem, second-year PhD student Dale Decatur and a team of researchers from assistant professor Rana Hanocka’s 3DL lab at the University of Chicago Department of Computer Science have created a technique called 3D Paintbrush.

PhD student Dale Decatur

“It turns out, that’s actually a pretty hard problem,” said Decatur, who developed the project. “What if we could just use text to describe some part and get an explicit segmentation of that part itself in 3D? That was how the first project started out.”

In the paper titled, “3D Paintbrush: Local Stylization of 3D Shapes with Cascaded Score Distillation,” the researchers introduce a new technique called Cascaded Score Distillation (CSD) that enables local, text-driven, high resolution editing. They observed that pre-trained 2D cascaded models — which the 3D model is built upon — possess different stages that each contain different levels of resolution and global understanding.

“Existing methods only use the first stage of the multi-stage supervision model,” Decatur stated. “I was actually surprised. Why not leverage the power of multiple stages? Because we use multiple stages, higher resolution stages have the ability to provide higher resolution supervision.” CSD, therefore, distills multiple stages to produce a high resolution local texture on an existing object.

Key to the method, the localization and stylization are learned simultaneously, interacting with each other to improve both the sharpness of the localization and the details of the edit. By employing this synergistic approach alongside CSD, 3D Paintbrush is able to achieve new levels of detail using different text prompts on a diverse group of 3D objects. When asked to place a Barcelona jersey on a Lego humanoid, 3D Paintbrush not only accurately locates the torso area to color the jersey, but also accurately synthesizes the crest, down to the golden football and the red “plus” sign. 3D Paintbrush is also not limited to “plausible” or realistic edits. Though no one has seen shin guards on a giraffe or a cow with a turtle shell, 3D Paintbrush can place both of those in their respective locations with a high degree of precision.

The official code is available on Github, where users can download the code along with a demo to see 3D Paintbrush in action. By inputting the text of a 3D object (e.g. “cow”) and the desired edit (e.g. “colorful crochet hat”), one can see how the computer infers the region without any manual selection by the user. As current changes to 3D models and avatars rely on manual region selection and editing by someone well-versed in 3D modeling tools, Decatur hopes to see this new technology applied to 3D modeling and video game animations. With 3D Paintbrush, the process would not just be entirely automated, but would also require no prior knowledge of 3D modeling software or specific expertise in the field, allowing for greater accessibility for novice or hobbyists to experiment in an area with growing potential.

“We hope that this research will allow everyone,” Decatur said, “regardless of technical background, to participate in the 3D modeling world.”

Building upon the 3D Paintbrush, Decatur plans to explore localized editing of the object’s geometry as well (e.g. adding horns on a cow) since currently 3D Paintbrush’s edits are constrained to the surface of the object. To learn more about his and the rest of 3DL’s research, you can visit their lab page.

Related News

More UChicago CS stories from this research area.
No Name

Prof. Rebecca Willett awarded the SIAG DATA Career prize

Feb 21, 2024
No Name

Argonne scientists use AI to identify new materials for carbon capture

Feb 19, 2024
No Name

Alumni Spotlight: Dixin Tang, Assistant Professor of Computer Science at UT Austin

Feb 05, 2024
No Name

NetMicroscope Uses AI to Improve Network Monitoring for a Better Internet Experience

Feb 01, 2024
No Name

NeurIPS 2023 Award-winning paper by DSI Faculty Bo Li, DecodingTrust, provides a comprehensive framework for assessing trustworthiness of GPT models

Feb 01, 2024
No Name

New research unites quantum engineering and artificial intelligence

Jan 29, 2024

“Machine Learning Foundations Accelerate Innovation and Promote Trustworthiness” by Rebecca Willett

Jan 26, 2024

Nightshade: Data Poisoning to Fight Generative AI with Ben Zhao

Jan 23, 2024
No Name

Group From UChicago CS To Present Four Papers at Most Prestigious International Quantum Conference

Jan 09, 2024
No Name

Alumni Spotlight: Get To Know Emily Wenger, a 2023 CS Graduate Who Was Just Named To The Forbes 30 Under 30 List

Nov 29, 2023
No Name

Three UChicago PhD Students From The Department of Computer Science Named To Forbes 30 Under 30 List

Nov 28, 2023
No Name

High School Students In The Collegiate Scholars Program Get To Know Robots

Nov 14, 2023
arrow-down-largearrow-left-largearrow-right-large-greyarrow-right-large-yellowarrow-right-largearrow-right-smallbutton-arrowclosedocumentfacebookfacet-arrow-down-whitefacet-arrow-downPage 1CheckedCheckedicon-apple-t5backgroundLayer 1icon-google-t5icon-office365-t5icon-outlook-t5backgroundLayer 1icon-outlookcom-t5backgroundLayer 1icon-yahoo-t5backgroundLayer 1internal-yellowinternalintranetlinkedinlinkoutpauseplaypresentationsearch-bluesearchshareslider-arrow-nextslider-arrow-prevtwittervideoyoutube