<ul><li>Video diffusion models (VDMs) facilitate the generation of high-quality videos.</li><li>Most research efforts in VDMs have focused on scaling during training, but scaling during inference time has received less attention.</li><li>Recent findings suggest that guiding the scaling inference-time search of VDMs with reward signals can enhance video quality by identifying better noise candidates.</li><li>The proposed ScalingNoise strategy improves global content consistency and visual diversity in long video generation.</li></ul>

ScalingNoise: Scaling Inference-Time Search for Generating Infinite Videos

Discover more