Yushu Wu

Ph.D. Candidate @ Northeastern University

I am a final year PhD student in Computer Engineering at Northeastern University, under the supervision of Prof. Yanzhi Wang. My research interests focus on Efficient Generative AI, Model Compression, Hardware-Software Co-design for Efficient AI.

I will be on job market since 2025 Fall (can graduate in 2025 Winter or 2026), and am actively looking for full-time research scientist or ML engineer. Please feel free to contact me via email (wu.yushu AT northeastern.edu) if our research match.

Google Scholar  /  LinkedIn  /  GitHub  /  OpenReview  /  Resume

Research

• Efficient Deep Learning: SnapGen-V (CVPR'25), Streamlined Inference for Video LDMs (NeurIPS'24), ToP-ViM (NeurIPS'24), Mobile- SR (ECCV'22), LOTUS (DAC'24), MOC (ICCAD'23).
• Generative AI: SnapGen-V (CVPR'25), Streamlined Inference (NeurIPS'24), SF-V (NeurIPS'24), ToP-ViM (NeurIPS'24).
• Efficient LLMs: ToR-SSM (EMNLP'24).

Recent News

Selected Publications

Google Scholar for all publications. * means equal contribution.

We propose a comprehensive acceleration framework to bring the power of the large-scale video diffusion model to the hands of edge users. We optimize current video diffusion model architecture for mobile deployment and conduct adversarial fine-tuning algorithm for our efficient model and reduce the denoising steps to 4. Our model, with only 0.6B parameters, can generate a 5-second video on an iPhone 16 PM within 5 seconds.

We present a novel, training-free framework named Streamlined Inference, which leverages the temporal and spatial properties of video diffusion models to significantly reduces peak memory and computational overhead, making it feasible to generate high-quality videos on a single consumer GPU.

We propose a novel approach to obtain single-step video generation models by leveraging adversarial training to fine-tune pre-trained video diffusion models. We show that, multi-step video diffusion model, can be trained to perform single forward pass to synthesize high-quality videos.

We revisit the unique computational characteristics of SSMs and discover that naive application disrupts the sequential token positions. This insight motivates us to design a novel and general token pruning method specifically for SSM-based vision models.

We identify the reasons for this failure and the limitations of current token reduction techniques on State Space Models. We propose a tailored, unified post-training token reduction method for SSMs.

To avoid thermal throttling and provide stable inference speed on the edge, we propose Lotus, a novel framework that is tailored for two-stage detectors to dynamicallyscale CPU and GPU frequencies jointly in an online manner based on deep reinforcement learning (DRL).

We identify a new class of neural network accelerators (e.g., NE16 on GAP9) that allows network weights to be quantized to different common (e.g., 8 bits) and uncommon bit-widths (e.g., 3 bits). Leveraging a differentiable NAS framework, the optimal bit-widths of a network is obtained for memory and energy efficient inference.

We propose DACO to reduce the power consumption of DNNs on mobile. A resource-oriented classifier is proposed to quantify the computation/memory intensity of DNN models and classify them accordingly. A set of rule-based policies is conducted for achieving the best-suited CPU-GPU system configuration.

We propose MOC, a Multi-Objective deep reinforcement learning-assisted DNN inference stage-adaptive CPU-GPU Co-optimization approach to reduce the power consumption of DNN inference meanwhile delivering an excellent speed on mobile devices.

We propose a compiler-aware SR neural architecture search (NAS) framework that conducts depth search and per-layer width search with adaptive SR blocks. A speed model incorporated with compiler optimizations is leveraged to predict the inference latency of the SR block with various width configurations for faster convergence.

We propose to use neural architecture search and network pruning to find a highly efficient network for image super-resolution. we are the first to achieve real-time SR inference (with only tens of milliseconds per frame) for implementing 720p resolution with competitive perceptual performance on mobile platforms.

Research Intern, Creative Vision, Snap Inc.

Worked on efficient text-to-video diffusion

Research Intern, Bose Corporation.

Worked on audio model quantization for NPU deployment