Adil Karjauv
Qualcomm AI Research. Amsterdam, Netherlands.
Hello, welcome to my personal page!
I am currently a Machine Learning Researcher (Senior Engineer) at Qualcomm AI Research working on efficient video generative AI. My work significantly contributed to the following:
- MobileWan: Led an architecture optimization effort for the largest video generation diffusion transformer running on-device (based on Wan2.2-5B), developed a novel structured pruning method, improving efficiency without sacrificing quality. Paper under review.
- PyramidalWan: Co-developed a step distillation method for a pyramidally-distilled video generation model, achieving a 10x inference speedup. Published at CVPR 2026; patent filed.
- Neodragon: Contributed to step distillation techniques (2.5-10x speedup) for the first diffusion transformer based mobile video generation model. Presented as a NeurIPS 2025 demo and published at ICLR 2026; patent to be filed.
- MoViE: Developed and optimized a state-of-the-art diffusion model for video editing, achieving an 87x on-device speedup – the fastest diffusion-based video editing demo shown on-device. Presented at NeurIPS 2024 and CVPR 2025 as a demo.
- Object-Centric Diffusion (OCD): Contributed to training-free techniques that reallocate compute toward salient regions in video editing, cutting latency up to 10x with comparable quality. Published at ECCV 2024.
- Pioneered the first on-device deep-learning-based video denoising solution at high resolution (QHD/4K at 30FPS); patent filed.
Prior to that, I completed a Master’s degree at KAIST in Robotics and Computer Vision (RCV) Lab under supervision of Prof. In So Kweon. My primary focus was on adversarial machine learning and its applications in multimedia. This work led to several publications in top-tier conferences and workshops such as CVPR, ICCV, NeurIPS, ICLR, and others.
I am always open to new opportunities and collaborations, so please feel free to reach out!
news
| Jul 07, 2026 | Our paper MobileWan: Closing the Quality Gap for Mobile Video Diffusion is now on arXiv! It’s the first 5B-parameter video diffusion model to run on a commercial mobile device, setting a new state-of-the-art for mobile video generation. |
|---|---|
| Jun 03, 2026 | Heading to CVPR 2026 in Denver — looking forward to connecting, come say hi! |
| Feb 20, 2026 | Our paper PyramidalWan: On Making Pretrained Video Model Pyramidal for Efficient Inference has been accepted at CVPR 2026! It presents a low-cost recipe for turning any pretrained video diffusion transformer model into a faster, pyramidal one — no retraining from scratch, no quality loss. |
| Jan 25, 2026 | Our paper Neodragon: Mobile Video Generation Using Diffusion Transformer has been accepted at ICLR 2026! It’s the first text-to-video generation model built specifically for mobile hardware, generating a 2-second video directly on-device in just ~6.7 seconds. |
| Dec 01, 2025 | Promoted to Senior Machine Learning Researcher at Qualcomm AI Research! |
selected publications
- ICCV 2021Data-free Universal Adversarial Perturbation and Black-box AttackIn Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021