Presentation

<
MultiAct: Text-to-Motion Generation from Composite Text via Tailored Attention Guidance
SessionMotion Generation, Warping & Control
DescriptionText-to-motion models struggle with composite prompts, often collapsing to a single action. We present MultiAct, an inference-time, unpaired framework that enhances pretrained generators by amplifying cross-attention for underrepresented tokens. A lightweight decision scheme selects optimal parameters per prompt, improving semantic coverage while preserving motion realism without retraining or architectural changes.
Authors
Nathan Sala
Tel Aviv University
Ofir Abramovich
Reichman University
Ariel Shamir
Reichman University
Daniel Cohen-Or
Tel Aviv University
Andreas Aristidou
University of Cyprus
CYENS Centre of Excellence
Sigal Raab
Tel Aviv University
Event Type
Technical Paper
TimeTuesday, 21 July 20262:40pm - 2:50pm PDT
LocationRoom 403 A
Digital Library PDF
Session Time & Location
Sunday, 19 July 20266:00pm - 8:45pm PDTHall K
Tuesday, 21 July 20262:00pm - 3:30pm PDTRoom 403 A
Keywords
Animation
Simulation
Registration Categories
Full Conference Supporter
Full Conference
Next PresentationNext PresentationAutoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation
Similar Presentations
Stylized Text-to-Motion Generation via Hypernetwork-Driven Low-Rank Adaptation
MOCHI: Motion Enhancement of Collaborative Human-object Interactions
Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation