BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CAIL//Events//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:CAIL Events
BEGIN:VEVENT
UID:2026-10-02-ishaan-chandratreya-and-david-charatan@cail.columbia.edu
DTSTAMP:20261013T005331Z
DTSTART:20261002T160000Z
DTEND:20261002T170000Z
SUMMARY:Vision Seminar: Ishaan Chandratreya and David Charatan - MilliVid:
  Hierarchical Latents for Long-Range Consistency in Video Generation
LOCATION:SCH 406
DESCRIPTION:Video generative models have become increasingly powerful\, bu
 t long-range consistency remains challenging to achieve because even a few
  dozen frames require impractically long transformer sequence lengths. We 
 show that this issue can be mitigated by generating video using coarse-to-
 fine rollout within a multi-scale token space. Our approach is simple: fir
 st\, we pre-train an autoencoder that compresses each frame into a hierarc
 hy of tokens\, with levels ranging from the typical latent resolution to o
 nly a handful of tokens per frame. The coarsest levels capture the most co
 nsequential information—such as scene layout and semantics—while finer
  levels add high-frequency appearance and texture. Then\, we train a video
  diffusion model to generate these tokens using coarse-to-fine rollout. By
  carefully controlling the level of detail at which frames are generated a
 nd used as context during each rollout step\, we are able to preserve long
 -range consistency in geometry and object permanence while spending less c
 ompute on the long-range consistency of less perceptually relevant details
 . We validate this approach using a custom dataset of long Minecraft video
 s\, where it produces substantially more consistent rollouts compared to e
 xisting baselines\n\nhttps://cail.columbia.edu/events/2026-10-02-ishaan-ch
 andratreya-and-david-charatan
URL:https://cail.columbia.edu/events/2026-10-02-ishaan-chandratreya-and-da
 vid-charatan
END:VEVENT
END:VCALENDAR