A scene needs a memory
Google’s September 24, 2026 research overview describes CANVAS, a system that plans continuity across generated shots. It keeps structured information about characters, locations and object states, alongside visual references. The useful idea is to carry a changing scene description forward instead of asking for each picture in isolation. [1]
What should stay, and what should change?
The April 15, 2026 CANVAS preprint describes a global plan followed by image generation that retrieves visual anchors and updates memory. A character’s appearance can change when the narrative calls for it; an object’s location must follow the events. Candidate images are checked against continuity questions before selection. [2]
A simple example
Hypothetical example: A traveler picks up a yellow suitcase beside a doorway. When the story returns to that doorway, the suitcase should be absent. Drawing the same suitcase twice is not enough: the sequence must account for who moved it. Our illustration shows this invented situation, not a CANVAS test result.
Go a little deeper
Optional reading · about 1 more minute
Continuity is not a complete physics model
The authors acknowledge expensive candidate generation, difficulty extracting entities from ambiguous narratives, and a small stress-test dataset. They also say the system does not fully model detailed physical interactions or complex motion across shots. Reported gains are the authors’ evaluation, not independent proof of reliable filmmaking. [2]
A useful way to inspect generated stories
Our interpretation: Look for a return visit. Does a room keep its layout after a cut elsewhere? Does an object retain the consequences of an earlier action? These questions reveal a different kind of failure from a visibly distorted single frame.
Original sources
Attributed synthesis, not original reporting. Examples labeled hypothetical or illustrative are explanatory. Reviewing a source does not independently validate its findings.
- Google Research: coherent long-form video generation ↗
September 24, 2026 developer overview, CANVAS subsection read September 29. Promotional perfect-coherence language and performance numbers not adopted.
- Mondal and colleagues: CANVAS ↗
April 15, 2026 v1 preprint; sections 3 and Limitations read September 29. Methods explanation, not independent replication.