Presentation

<
LGTM: Local-to-Global Text-driven Human Motion Diffusion Model
SessionCharacter Animation From Data
DescriptionWe introduce LGTM, a novel Local-to-Global pipeline for Text-to-Motion generation based on diffusion model. It decomposes motion description to body-part level with LLMs and encodes them with corresponding body part motion individually, then optimizes whole body motion by attention encoder. As a result, it can generate local semantics well-matched motion.
Authors
Haowen Sun
Shenzhen University
Ruikun Zheng
Shenzhen University
Haibin Huang
Kuaishou Technology
Chongyang Ma
ByteDance Inc.
Hui Huang
Shenzhen University
Ruizhen Hu
Shenzhen University
Event Type
Technical Paper
TimeWednesday, 31 July 202410:45am - 10:55am MDT
LocationMile High 4
ACM Digital Library Journal Papers' PDFs
Conference Papers' PDFs
Session Time & Location
Sunday, 28 July 20246:00pm - 8:45pm MDTBluebird Ballroom
Wednesday, 31 July 202410:45am - 12:15pm MDTMile High 4
Interest Areas
Research & Education
Recordings
Livestreamed
Recorded
Keywords
Animation
Registration Categories
Full Conference
Full Conference Supporter
Virtual Access
Exhibitor Full Conference
Wednesday
Next PresentationNext PresentationTaming Diffusion Probabilistic Models for Character Control