|
Dishani Lahiri I am a Machine Learning Engineer 2 at Snap Inc. in New York City, working on post-training efficient generative models for image and video. I completed my Master of Science in Computer Vision (MSCV) in the Robotics Institute at Carnegie Mellon University in Dec 2023, where I was a Research Assistant at KLab, advised by Prof. Kris Kitani, working on visual-inertial state estimation for Meta Aria glasses. During my summer internship at Slingshot AI, I worked on fine-tuning personalized text-to-image diffusion models and LLaMA2-7B for personalized text style transfer. Before that, I worked at Samsung R&D Institute, Bangalore, where I helped develop and deploy AI Night mode and the Expert RAW app for Samsung's flagship devices. I completed my undergraduate studies in ECE at DTU in 2019, advised by Prof. S. Indu and Prof. D.K. Vishwakarma. Email / Google Scholar / LinkedIn / Github |
|
Snap Inc. |
CMU |
KLab@CMU |
Slingshot AI |
Samsung R&D Institute |
DTU, Delhi |
Projects & PublicationsI'm interested in computer vision, natural language processing, and machine learning, especially in building personalized multi-modal solutions for edge devices. |
|
CVPR, 2026 paper | poster A streaming sandwich diffusion transformer for efficient, high-fidelity video generation on mobile devices. S2DiT uses a mixture of LinConv Hybrid Attention and Stride Self-Attention, a sandwich architecture found via budget-aware dynamic programming, and a 2-in-1 distillation framework from large teacher models, achieving quality on par with state-of-the-art server video models while streaming at over 10 FPS on an iPhone. |
|
arXiv preprint, 2026 paper A diffusion transformer for efficient, high-fidelity image generation on edge devices, combining an adaptive global-local sparse attention mechanism, an elastic training framework for dynamic adjustment across hardware targets, and a distillation pipeline enabling high-fidelity, low-latency (4-step) generation suitable for real-time on-device use. |
|
CVPR, 2025 (Highlight) paper | project page An efficient text-to-image diffusion model for mobile devices, combining architectural design choices, cross-architecture knowledge distillation from larger models, and adversarial few-step guidance. SnapGen generates 1024x1024 images on a mobile device in about 1.4s with only 379M parameters, outperforming significantly larger models on standard benchmarks. |
|
Dishani Lahiri*, Neeraj Panse*, Moneish Kumar* ICCV, 2023 Workshop on AI for 3D Content Creation paper | code | webpage We present our method for transferring style from any arbitrary image(s) to object(s) within a 3D scene. Our primary objective is to offer more control in 3D scene stylization, facilitating the creation of customizable and stylized scene images from arbitrary viewpoints. To achieve this, we propose a novel approach that incorporates nearest neighborhood-based loss, allowing for flexible 3D scene reconstruction while effectively capturing intricate style details and ensuring multi-view consistency. |
|
Dishani Lahiri*, Chhavi Dhiman, Dinesh Kumar Vishwakarma IEEE, 2017 Conference on Information and Communication Technology (CICT) paper We propose a solution to detect abnormal human actions in the image using Histogram of Oriented Gradients (HoG) as the feature descriptor, Principal Component Analysis (PCA) as the dimensionality-reduction technique, and Support Vector Machine as the ML tool for classification. We also release a dataset for abnormal human activities of fainting, headache, and chest pain. |
|
|