|
MobileWan: Closing the Quality Gap for Mobile Video Diffusion |
Mohsen Ghafoorian*, Denis Korzhenkov*, Adil Karjauv*, Ioannis Lelekas*, Noor Fathima*, Spyridon Stasis, Hanno Ackermann, Boris van Breugel, Markus Nagel, Fatih Porikli, Animesh Karnewar, Amirhossein Habibian
* Core equal contribution ## Citation ```bibtex @article{ghafoorian2026mobilewan, title = {MobileWan: Closing the Quality Gap for Mobile Video Diffusion}, author = {Mohsen Ghafoorian and Denis Korzhenkov and Adil Karjauv and Ioannis Lelekas and Noor Fathima and Spyridon Stasis and Hanno Ackermann and Boris van Breugel and Markus Nagel and Fatih Porikli and Animesh Karnewar and Amirhossein Habibian}, journal = {arXiv preprint arXiv:2607.06173}, year = {2026} } ``` ## Overview MobileWan starts from the [Wan-AI/Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) text-to-video transformer and modifies it to make high-quality video diffusion more practical for mobile deployment. This checkpoint focuses on reducing the computational cost of the transformer while largely preserving generation quality. The main transformer-side changes include: - **Attention-head pruning:** We prune less important attention heads in the Wan2.2 5B transformer to reduce compute and memory costs while retaining the most useful attention pathways. This checkpoint specifically prunes ~23% of the heads. - **Recurrent Hybrid Attention:** We use recurrent hybrid attention to make the transformer more efficient for video generation, reducing the cost of long spatio-temporal attention patterns with linear attention while keeping the more crucial local dependencies modeled through softmax attention. - **Step distillation:** We reduce the number of diffusion steps required at inference time. For this checkpoint, the step-distillation stage uses the decoupled DMD objective, and our evaluations use sampling with 3 denoising steps. For more details, please read the corresponding paper: [https://arxiv.org/pdf/2607.06173](https://arxiv.org/pdf/2607.06173). ## How to Sample Videos Please refer to: [https://github.com/qualcomm-ai-research/mobilewan](https://github.com/qualcomm-ai-research/mobilewan) ## Model Description - **Developed by:** Qualcomm AI Research, Generative Vision group, Amsterdam, Netherlands - **Model type:** Mobile video generation with efficient diffusion transformer architecture - **Model size:** ~5B parameters (DiT only) - **Model precision:** torch.bfloat16 (BF16) - **Model resolution:** This model is developed to generate 81-frame videos (5s @ 16fps) at [480 x 832] resolution directly on a Snapdragon-powered mobile phone. - **Base model:** [Wan-AI/Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) - **Pipeline tag:** Text-to-video - **Description:** This model can be used to generate videos based on provided text prompts. Note that the optimized video decoder is not released here; therefore, the sampling code uses the decoder from the original Wan2.2 5B pipeline. - **Resources for more information:** Check out the [GitHub repository](https://github.com/qualcomm-ai-research/mobilewan) and the [technical report on arXiv](https://arxiv.org/abs/2607.06173). ## License/Terms of Use This model is released under the BSD 3-Clause Clear license and the Qualcomm responsible AI license: https://www.qualcomm.com/site/responsible-ai-license ## Uses The model is intended for research purposes. Possible research areas and tasks include: - Research on efficient transformer or non-transformer based backbone architectures for video generation. - Generation of video-based artworks and use in design and other artistic processes. - Applications in educational or creative tools. - Research on generative models. - Safe deployment of models which have the potential to generate harmful content. - Probing and understanding the limitations and biases of generative models. ## Limitations and Bias ### Limitations - The model does not achieve perfect photorealism. - The model cannot render complex legible text. - The model's generation quality may be impacted by prompts that are too short. - The model cannot produce videos with accurate physically compliant motion. ### Bias While the capabilities of the presented mobile video generation model are impressive, they can also reinforce or exacerbate social biases inherited from the base Wan model.