Skip to content

SYSTEM Cited by 1 source

CANVAS

CANVAS (Continuity-Aware Narratives via Visual Agentic Storyboarding) is a Google Research multi-agent framework that explicitly plans visual continuity in multi-shot narratives (to appear at EMNLP 2026). Built as an orchestration layer on top of Gemini, it addresses character drift and unstable environments that arise even when the script is unified. (Source: sources/2026-09-24-google-automating-coherent-long-form-video-generation)

Core mechanism: persistent visual memory (world-state tracking)

CANVAS enforces coherence by maintaining structured representations of characters, locations, and object states as the narrative evolves — an explicit world-state model. It relies on a persistent visual memory:

  • Retrieve visual anchors from memory when a setting/character recurs, or initialize new anchors when needed → smooth transitions within a setting.
  • Ensures characters retain identity and environments preserve spatial structure when revisited — handling both consecutive transitions (a character's clothing shot-to-shot) and non-consecutive transitions (the camera returns to the main hall after a detour).

This is agent memory specialized to visual world-state: the memory substrate is the mechanism that prevents drift, the same way tenant/session memory prevents context loss in text agents.

Why it matters

Unguided generation drifts. In a multi-shot museum-heist evaluation:

  • Gemini-3.1-Pro alone exhibits prop inconsistency (the artifact changes) and background drift (room layout shifts).
  • AutoStudio (an alternative multi-agent baseline) degrades across cuts — character drift (the thief's cap disappears) + background inconsistency.
  • CANVAS keeps characters, spatial geometry, and object states coherent across the entire narrative via its persistent visual memory.

Results

Continuity gains across scene reappearances on ST-Bench and HardContinuityBench (the latter built from GPT-5.2 multi-shot storyboards with massive reappearance gaps, frequent costume/accessory changes, and complex prop state changes).

Seen in

Last updated · 766 distilled / 2,225 read