Reviewing AI video with separate checks for motion, subject, and background via VL models

professr_dumbledore · reddit · 2026-10-03

A concrete workflow for reviewing AI-generated video with a vision-language model, using Cosmos3-Edge for generation and Ling 3.0 VL reading six sampled frames against the intended motion.

The example runs two generation instructions — an initial climb-then-right plan, then a Take B with hover 0–5s and drift right 5–10s — with cyan guides at fixed coordinates comparing start and end frames.
Core method: split image-to-video review into three separate questions:
Motion: where is the subject at start and end, on the same reference axes?
Subject: which identifying details remain visible across samples?
Background: how do background landmarks line up frame to frame?
Separate questions make the next prompt revision specific: direction maps to the path, subject to appearance, background to the scene.

Original post →

Reviewing AI video with separate checks for motion, subject, and background via VL models

More from Multimodal