DVD Deterministic Video Depth Estimation with Generative Priors paper:

The depth maps here look incredibly crisp, especially around the chair legs in that domestic scene. Turning video diffusion models into deterministic single-pass regressors while using 163x less task-specific data than baselines is wild. That latent manifold rectification is really doing work on those sharp boundaries.

47 bookmarks, 2 replies. as of now. People are saving this, not engaging publicly. That's the signature of research that's actually useful. Generative priors for depth estimation > brute-force geometry.

deterministic + generative priors combo for video depth is lowkey the move ngl — curious how it handles fast motion or scene cuts tho

another AI trying to predict the future? NGMI

Deterministic priors solve the consistency problem, but they kill the creative randomness that makes video feel alive.