Title: From Zero to Hero: Training-Free Custom Concept Spawning in World Models

Abstract

Autoregressive world models have established themselves as a robust framework for interactive video generation, enabling users to explore dynamically created environments through their own actions. Typically, these systems rely on a text prompt and/or a single reference frame to generate the entire scene. However, a critical flaw emerges when users move beyond the boundaries of that initial frame: the unseen areas are filled based solely on the base model’s inherent priors, leaving users with no control over what appears or where it is placed. This lack of controllable scene composition represents a significant barrier for applications like gaming, interactive storytelling, and simulation.

We define this missing functionality as "concept spawning"—the ability to introduce a user-defined visual concept into a world model, mirroring the mechanics of spawning objects in a game engine. To address this, we present SPAWN (Swapping Pinned Anchor with Windowed iNjection), a novel, training-free approach to concept spawning. SPAWN leverages a specific structural characteristic of image-to-video backbones: the first slot in the context memory is fixed to the reference frame, serving as a foundational anchor for all generated segments.

By temporarily replacing this anchor with an external concept latent within a brief injection window—and subsequently restoring the original anchor—we enable the concept to propagate organically throughout the rollout via the model’s internal memory. SPAWN is versatile, accommodating inputs ranging from fine-grained details like characters and props to large-scale structures such as buildings and landmarks. It accepts either a concept image or a text description as input. Our experiments demonstrate that SPAWN successfully integrates concepts with consistent lighting, scale, and perspective, while maintaining identity and temporal coherence. These results confirm that controllable concept spawning is feasible within existing autoregressive world models without the need for additional training.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC