CineMobile targets on-device image-to-video generation for cinematic camera effects such as bullet time, dolly zoom, and slow motion. The method combines distillation-guided pruning, diffusion distillation, reinforcement learning, and hybrid post-training quantization. The resulting generator uses four denoising steps and has a model footprint below 1 GB. According to the paper abstract, it is 40x faster than a Wan 2.1-based teacher while retaining comparable visual quality. It generates 49-frame, 480p videos with 0.6 seconds per denoising step on an NVIDIA H200 and 20 seconds per step on a MediaTek Dimensity 8400 Ultimate platform, using 1.8 GB peak memory.
No heat snapshots are available in the last 24 hours.