AI

COSMI: COmpositional Synthesis of Multi-object Interactions

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 17:18:24 UTC.

Briefings

COSMI generates a dataset of 222k sequences and 275 hours of video featuring up to five objects, creating a text-to-interaction diffusion transformer that learns from single-object interactions to predict multi-object contact-consistent clips. The method generalizes to unseen object and interaction combinations, outperforming baselines in text alignment and contact accuracy, with the largest margin on the unseen object. The dataset is publicly available via a project page.

Key points

• COSMI dataset contains 222k sequences and 275 hours of video with up to five objects.
• COSMI method uses a text-to-interaction diffusion transformer trained on single-object interactions.

Context

Generative models of human-object interaction are limited by multi-object capture costs; COSMI overcomes this by composing single-object clips that balance hands, bodies, and language model checks to ensure plausible pairings.

Limitations

Evidence is limited to the COSMI dataset pipeline; no additional background on the method's training data or evaluation metrics is supplied.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 22/10/2026, 14:45:21 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigrwrekpy5kghctrsmqbsofchj4nqsel7zgpvypr3o7oy2ss4nuam