Qualcomm Neodragon: Mobile Video Generation Using Diffusion Transformer
Qualcomm has developed Neodragon, a novel text-to-video generation system optimized for mobile hardware. This system can produce short videos in under 7 seconds directly on a Qualcomm Hexagon NPU, achieving a VBench score of 81.61. Neodragon employs several optimization techniques, including a distilled text encoder, an asymmetric decoder distillation, and pruning of diffusion model blocks, to reduce model size and computational requirements while maintaining high fidelity. The goal is to enable on-device, private, and cost-effective AI-based video content creation. AI
IMPACT Enables on-device, private, and cost-effective AI-based video content creation for mobile users.