AI Generates Realistic Fluid Videos with Physics Grounding

Current video diffusion models often fail to accurately represent fluid dynamics, producing visually appealing but physically incorrect content where liquids defy gravity or momentum. This limitation is attributed to a lack of explicit motion supervision in large-scale video-text datasets, leading models to mimic appearance rather than underlying physics. Researchers have addressed this by introducing two key contributions. First, they developed a novel physics-simulation fluid dataset, combining over 1,600 simulated pouring and sloshing videos with thousands of real pouring videos. This dataset provides a rich source of physically accurate motion data. Second, they implemented a dual-stream image-to-video architecture built upon a pre-trained diffusion-transformer. This architecture augments the standard RGB decoder with a lightweight Optical-Flow Decoder branch, explicitly trained with end-point-error and smoothness losses. This optical flow stream, fused into the RGB stream, allows the model to internalize a coherent motion prior. The method significantly improves physical commonsense and video quality scores, outperforming existing solutions and being preferred by human raters.