Researchers have developed COSMI, a novel method for synthesizing multi-object interactions in generative models. This approach addresses the limitations of existing datasets by composing single-object interaction clips into more complex scenarios, significantly expanding the available data. The COSMI dataset contains 222,000 sequences and 275 hours of data, featuring up to five objects, which is nearly thirty times larger than previous multi-object datasets. A text-to-interaction diffusion transformer model trained on this data demonstrates strong generalization capabilities on unseen objects and interaction combinations, outperforming baseline methods in text alignment and contact accuracy. AI
IMPACT Enables more complex and realistic human-object interactions in generative AI, potentially improving applications in robotics and virtual environments.
RANK_REASON The cluster describes a new dataset and method for generative models, presented in a research paper.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
