Van Video Models

3D VAE (Variational Autoencoder) is a spatial-temporal compression technology that encodes video data into a compact latent space. By processing spatial details and temporal motion simultaneously, it preserves cinematic textures and fluid dynamics while significantly reducing computational overhead.

Flow Matching is a state-of-the-art generative framework that defines a continuous, straight-line path between noise and data. Compared to traditional diffusion models, it enables more precise motion control and faster convergence, resulting in high-fidelity video with complex physical logic.

While built on the Wan 2.5 and 2.6 architectures, Van Model is an optimized flagship series that delivers significantly higher output resolutions and greater creative freedom. Through proprietary distillation, Van also offers drastically faster inference speeds and lower operational costs.

Van optimizes the 3D VAE decoder and integrates high-fidelity refinement layers during the compute distillation process. This ensures that while the core movement and logic remain identical to the base model, the visual clarity and pixel density are substantially enhanced.

Compute distillation compresses complex model knowledge into a highly efficient inference engine. This allows Van to generate high-quality video using fewer sampling steps, shattering the "quality-to-cost" barrier and enabling lightning-fast production for large-scale workflows.