Researchers have compressed a video diffusion model small enough to run locally on a midrange phone chip.
CineMobile starts with Wan 2.1, a large Diffusion Transformer, and applies three layers of compression: structural pruning guided by a distillation process, a four-step denoising schedule trained with a mix of diffusion distillation and reinforcement learning, and post-training quantization that pushes the final model below 1 GB. The result generates 49-frame 480p clips — think bullet-time or dolly-zoom effects — on a MediaTek Dimensity 8400 Ultimate 5G chip. Each of the four denoising steps takes around 20 seconds on that chip, putting total generation time at roughly 80 seconds. The paper's headline 40x speedup figure is a comparison against the original Wan 2.1 teacher model on the same hardware — not a mobile-versus-cloud benchmark. Large Diffusion Transformers on mobile silicon would otherwise take many minutes per video, so the gap is plausible even though the paper does not state the teacher's mobile latency explicitly.
Most on-device video generation today means a cloud API call or a heavily degraded local effect — both carry latency, bandwidth costs, or quality penalties. A sub-1 GB model that completes in under two minutes on a chip already shipping in midrange Android phones changes that equation, at least on paper. Privacy-sensitive use cases have an obvious reason to prefer inference that never leaves the device.
This is still a preprint, and a working research demo is not a shipping product. Getting 1.8 GB peak memory, 80-second generation time, and usable cinematic quality to coexist inside a real app is the harder problem the paper does not solve.