ByteDance Seed and academic partners introduce Trace Anything, a feed-forward 4D trajectory-field model

Sign in✓ Signed in
Researchers from ByteDance Seed, HKUST, Zhejiang University, and Dartmouth College have released Trace Anything, a neural network that represents any video as a dense field of continuous, parametric 3D trajectories—one B-spline path per pixel per frame—predicted in a single feed-forward pass. The work, posted on arXiv on 15 October 2025 and listed as an ICLR 2026 paper, is accompanied by official code, a project page, and model weights.
Continue reading
The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in✓ Signed in — this article isn’t included in your current plan.
Unlocking the full article...