GroundShot: Visually Consistent Multi-Shot Long Video Generation via Entity-Grounded Shot Scheduling

2026-06-18

AI Summary

This work addresses the challenge of appearance drift in multi-shot long video generation, where cross-shot entities often exhibit visual inconsistencies. The authors propose a training-free, model-agnostic agent framework that, for the first time, defines entity consistency from the viewer’s perceptual perspective. By constructing an entity-level visual memory online and integrating an entity anchoring mechanism with a dynamic reference scheduling strategy, the method retrieves high-fidelity references and optimizes shot generation order prior to synthesis. Notably, it enhances multi-shot consistency without modifying the underlying generative model. The approach is evaluated on GroundBench, a newly introduced entity-level benchmark, demonstrating superior performance over existing methods while preserving the base model’s native capabilities.

Abstract

Generating visually consistent multi-shot videos remains an open challenge. As videos span more shots, inconsistencies can accumulate across shots, causing entities that reappear across shots -- characters, objects, and locations -- to drift away from how they first appear. We observe that viewers judge consistency by comparing each later appearance of an entity with its first clear appearance; the visual quality of this initial appearance sets the consistency ceiling for all that follows. Motivated by this, we present \textbf{GroundShot}, a training-free, model-agnostic agentic framework for entity-grounded multi-shot generation. GroundShot builds an entity-level visual memory online from accepted generated shots: it schedules shots' generation order by their expected usefulness as entity references, grounds entities from generated videos, verifies their reliability before adding them to memory, and retrieves suitable entity references from memory before each shot is generated. To evaluate this entity-centered view of consistency, we further introduce \textbf{GroundBench}, a diagnostic benchmark that measures consistency at the entity level while isolating controlled challenge dimensions. Experiments show that GroundShot improves multi-shot consistency over existing methods while requiring no additional training or model modification.

Problem

• visually consistent
• multi-shot video generation
• entity consistency
• video generation
• temporal consistency

Innovation

• entity-grounded generation
• multi-shot video consistency
• visual memory
• shot scheduling
• training-free framework