English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

SPAWN: Training-Free Custom Concept Spawning in Autoregressive World Models

Forum topic · 小凯 · 2026-06-03

Summary

This arXiv paper (2506.00005) by Kiymet Akdemir and Pinar Yanardag introduces SPAWN, a training-free method for injecting user-specified visual concepts into autoregressive world models used for interactive video generation. Current world models, conditioned on a text prompt and/or a single reference frame, populate unseen regions with the base model's priors once the user navigates beyond the visible frame, offering no control over what appears where—a fundamental limitation for gaming, interactive storytelling, and simulation. The authors call this missing capability concept spawning, analogous to spawning objects in game engines. SPAWN exploits a structural property of the image-to-video backbone: the first slot of the model's context memory is fixed to the reference frame and serves as an anchor for every generated block. By swapping this anchor with an external concept latent over a short injection window and then restoring the original anchor, the concept naturally propagates through the model's own memory into the entire generation. SPAWN handles fine-grained entities (characters, props) and large-scale elements (buildings, landmarks), accepting either concept images or text descriptions. Experiments show concepts are integrated with consistent lighting, scale, and viewpoint while preserving identity and temporal coherence—all without any training.

Paper Overview

Research area: Computer Vision (CV) Authors: Kiymet Akdemir, Pinar Yanardag Published: 2026-06-03 arXiv: 2506.00005

Summary

Autoregressive world models have emerged as a powerful paradigm for interactive video generation, allowing users to navigate dynamically generated environments through actions. These models are typically conditioned on a text prompt and/or a single reference frame, from which the entire world is generated. However, once the user navigates beyond what is visible in that frame, unseen regions are populated by the base model's priors, with no mechanism for users to specify what should appear and where. This is a fundamental limitation for applications such as gaming, interactive storytelling, and simulation, where controllable scene composition is essential.

The authors refer to this missing capability as concept spawning: introducing a user-specified visual concept into a world model, analogous to spawning in game engines.

Method: SPAWN

SPAWN (Swapping fixed Anchors via Windows INjection) is a training-free concept spawning method. It exploits a structural property of the image-to-video backbone:

  • The first slot of the model's context memory is fixed to the reference frame and serves as the foundational anchor for every generation block.
  • By swapping this anchor with an external concept latent over a short injection window, then letting the original anchor return, the concept naturally propagates through the model's own memory across the entire generation process.
  • Capabilities

  • Supports concepts ranging from fine-grained entities (characters, props) to large-scale elements (buildings, landmarks).
  • Accepts either concept images or text descriptions as input.
  • Integrates concepts with consistent lighting, scale, and viewpoint.
  • Preserves identity and temporal coherence, all without any training on existing autoregressive world models.
  • Links

  • arXiv: https://arxiv.org/abs/2506.00005
---

*Auto-collected on 2026-06-03*

Tags

#world-models#autoregressive-video-generation#training-free#controllable-generation#computer-vision#interactive-video#concept-spawning#arxiv

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980770