Research map / Video / Video generation
From Zero to Hero: Training-Free Custom Concept Spawning in World Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SPAWN, a training-free method that swaps the pinned first context-memory anchor of an autoregressive image-to-video world model (WorldPlay) with a user concept latent for a short window, so the concept propagates through the rollout.
Problem
Autoregressive interactive world models generate unseen regions from priors, giving users no way to specify which visual concept appears where during navigation (concept spawning).
Approach
Exploits the first context-memory slot, which is pinned to the reference frame and acts as an attention-sink-like anchor; replacing it with an external concept latent over a windowed set of chunks lets the concept enter the temporal and spatial KV memory and persist after the original anchor is restored, with no training.
Key results
- SPAWN reaches the highest VBench Overall score of 0.901, versus 0.846 for WorldPlay, 0.853 for Wan 2.2 and 0.749 for HunyuanVideo.
- In the user study, SPAWN gets prompt-following 3.961 vs 2.393 for the next-best method (Wan 2.2), and motion 4.587 vs 4.438 for WorldPlay.
- Concept-conditioned I2V Subject/Background reach 0.885/0.907 with 5 injection chunks, rising from 0.793/0.801 at one chunk.
Surprising findings
- Replacing all anchor slots gives the highest raw concept similarity (I2VSC 0.963) but collapses scene fidelity (I2V Subject 0.852).
- Injecting only the last slot or a middle slot fails to spawn the concept properly, while replacing slot 0 alone or slots 0+1 works.
Limitations
- Inherits the camera trajectories and action space of the base world model; cannot spawn concepts outside the reachable region.
- Evaluated only on WorldPlay and on 8-second rollouts for fair comparison with I2V baselines.
- Concept-conditioned metrics are self-referential, and baselines lack concept-image input, so direct comparison is limited.
Experimental setup
Datasets: WorldPlay scene references (panoramas)
Benchmarks: VBench VBench-I2V
Metrics: Aesthetic Quality Imaging Quality Background Consistency Dynamic Degree Motion Smoothness Subject Consistency Camera Motion I2V Subject I2V Background I2V Subject (Concept)
Baselines: HunyuanVideo Wan 2.2 WorldPlay
Builds on: WorldPlay HunyuanVideo PRoPE Self-Forcing attention sink
Closest papers by meaning
- Genie: Generative Interactive Environments (ICML 2024 · similarity 0.88)
- Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.87)
- Unbounded: A Generative Infinite Game of Character Life Simulation (ICLR 2025 · similarity 0.87)
- WorldMem: Long-term Consistent World Simulation with Memory (NeurIPS 2025 · similarity 0.87)
- Learning Interactive Real-World Simulators (ICLR 2024 · similarity 0.87)
- GameGen-X: Interactive Open-world Game Video Generation (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.