We present a demo of our Neodragon Mobile VDM as a plugin inside Adobe Premiere Pro. This plugin is being run on a laptop device which has a Qualcomm Snapdragon X Elite SoC, and containing the same Hexagon NPU that is present in Mobile devices running on Snapdragon 8 Gen 2 SoC. As shown by TaskManager, the text-to-video model runs on the Hexagon NPU, and the entire video generation pipeline runs on device to generate the video clip. This Plugin demonstrates how our Mobile VDM can be easily integrated into existing video editing software, therby accelerating various creative workflows.

Our optimizations enable high-quality video generation on mobile devices. The results demonstrate that both fidelity and diversity are preserved, with no compromise due to our optimisations. Our model achieves a new state-of-the-art in mobile text-to-video generation, reaching a vbench score of 81.61. The prompts that generated the videos are available in the filenames of each video.

Video results for the qualitative evaluation of our proposed Text-Encoder Distillation aproach. Note that The SSD-1B first frame generator is not present for these results, and the generated videos are at native (non super-resolved) resolution of [49 x 320 x 512].

Video results for the qualitative evaluation of our proposed Asymmetric Decoder Distillation aproach. Note that The SSD-1B first frame generator is not present for these results, and the generated videos are at native (non super-resolved) resolution of [49 x 320 x 512]. As can be noticed, there is no perceptible difference in quality between the two decoders.

Video results for the qualitative evaluation of Stage-1 finetuning of Block Pruning. This shows the effectiveness of our block selection for pruning approach. Although the technique is simple, but it can be very effectively applied to obtain a control over the trade-off between quality v/s model size for the MMDiT denoiser.

Video results for the qualitative evaluation of Stage-2 finetuning of Block Pruning. We show how much qualitative difference applying Stage-1 finetuning and followed by Stage-2 finetuning to the 18-blocks pruned model. As can be seen, the quality improves significantly after Stage-2 finetuning enabling near loss-less model compression.

Video results for the qualitative evaluation of our Step Distillation experiments. As can be seen, PYramidal-DMD preserves the best motion dynamics among all the distillation methods, although it introduces colour saturation and semantic artifacts. We note that Pyramidal-Mean flows also shows some promising results, but because it doesn't get as high a VBench as Pyramidal-DMD, we choose DMD as our final approach. This is the reason why, we replace first frame to the one generated from SSD-1B.