Render, Don't Decode: Weight-Space World Models with Latent Structural Disentanglement

Researchers introduce NOVA, a world modeling framework that represents scene state as weights in implicit neural representations (INRs) rather than traditional encoded latent spaces. The approach eliminates decoder bottlenecks, achieves structural disentanglement of scene components, and enables controllable video generation on consumer GPUs with only 40M parameters.

NOVA represents a meaningful shift in how AI systems approach world modeling by replacing the conventional encode-latent-decode pipeline with a weight-space representation. This architectural innovation directly addresses computational inefficiency that has plagued video understanding systems, which typically require heavy decoder networks to reconstruct frames from compressed representations. By storing scene information as the parameters of coordinate-based neural networks, the framework achieves remarkable compactness while maintaining interpretability—a significant advantage over opaque latent spaces that resist analysis.

The technical innovation builds on growing interest in implicit neural representations and neural rendering as alternatives to explicit pixel-space approaches. This trend reflects broader recognition that traditional autoencoders waste capacity on reconstruction when downstream tasks only require understanding dynamics and structure. NOVA's ability to automatically disentangle background, foreground, and motion without explicit supervision demonstrates that structured weight-space representations naturally encourage interpretable feature learning.

The framework's implications extend across AI development and resource efficiency. Operating at 40M parameters on consumer hardware makes world modeling accessible to smaller labs and researchers, potentially democratizing advanced video understanding research. The zero-shot super-resolution capability and content-editing features without compromising dynamics suggest practical applications in synthetic media generation and robotics training.

For the AI research community, this work validates the hypothesis that representation structure matters as much as raw model capacity. Future development will likely focus on scaling these approaches to longer sequences, higher resolutions, and more complex multi-object scenes. The rendering-based paradigm may influence how other generative models fundamentally approach state representation.

• →NOVA uses implicit neural representation weights as world model state, eliminating heavy decoder networks and improving efficiency
• →The framework achieves automatic disentanglement of scene components without auxiliary losses or adversarial training
• →Model operates efficiently at 40M parameters on single consumer GPU, democratizing world model research
• →Rendering-based approach enables zero-shot super-resolution and independent editing of content versus dynamics
• →Weight-space representations demonstrate superior interpretability compared to traditional opaque latent spaces