According to a Synced Review report, Adobe Research has proposed a new approach to video world models that addresses a persistent problem: long-term memory. The researchers combine State-Space Models (SSMs), which are efficient at modeling long-range dependencies, with dense local attention to maintain coherence across frames.
The design also relies on specialized training strategies, including diffusion forcing and frame local attention. These techniques help the model learn both the global structure of a video sequence and the fine-grained local details needed for consistent generation.
Because only one source was reviewed, there are no differing accounts to compare. The report frames the work as overcoming a long-standing limitation in video world models, though independent validation and further details would be needed to assess the full scope of the results.