Diverse Score Distillation

Yanbo Xu Jayanth Srinivasa Gaowen Liu Shubham Tulsiani · 2024

Read Diverse Score Distillation by Yanbo Xu Jayanth Srinivasa Gaowen Liu Shubham Tulsiani (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-06780.txt.

Description

Yanbo Xu and co-authors present a method that makes image and 3D generation more varied, in work from 2024. A common tool guides 3D building with scores from a 2D image model trained to produce pictures from text. That tool tends to give near-identical results, even though the image model itself can produce varied pictures.

The new method borrows from how the image model samples pictures step by step. It guides building along paths set by random starting seeds, which spreads the results out. The team adjusts the idea for cases where the building steps cannot follow those paths exactly, such as 3D shapes whose views change together. Tests on 2D tasks, text-based 3D building, and rebuilding a scene from one photo show more varied results with quality kept.

Catalog record

CREATORYanbo Xu Jayanth Srinivasa Gaowen Liu Shubham Tulsiani
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-06780

Browse the full library · About the Tower of Babel