Export
XING WANG
· 2 previous affiliations
- ByteDance (China)Since 2019
- Engineer · Since 2019
- Simon Fraser University2012–2017
- PhD student · 2012–2017
6 other authors named XING WANG: Xing Wang (TikTok · ByteDance (China) · 81 papers), Xing Wang (China Mobile Research Institute · 23 papers), Xing Wang (7 papers), Xing Wang (4 papers), Xing Wang (2 papers), and 1 more
Research Overview
Xing Wang focuses on efficient neural network architectures, hardware-aware model design, and the acceleration of generative models. His early work explored neural architecture search and vision transformers optimized for edge deployment. More recently, he investigates distillation techniques, quantization, and caching strategies to improve the inference efficiency of diffusion models and autoregressive video generation.
Research Trajectory
Wang's research trajectory begins with a focus on optimizing neural network architectures for efficient deployment. In 2022, he contributed to the research direction of Mitigating Weight Interference in Weight-Sharing Architecture Search through his collaborative work on Progressive Automatic Design of Search Space for One-Shot Neural Architecture Search. This work addressed the discrepancy between one-shot model accuracy and stand-alone training performance by proposing a progressive search strategy based on operation pruning to build a layer-wise operation search space.
During this period, Wang also co-developed novel vision transformer (ViT) architectures designed to balance computational cost with predictive performance. He contributed to SepViT: Separable Vision Transformer, which decoupled self-attention into local depthwise and global pointwise operations. This approach advanced the broader area of Linear Complexity Attention via Dimensional Decomposition and addressed the research area of Why Theoretical FLOPs Fail to Predict Real Latency by significantly reducing inference latency compared to similarly accurate models. Extending this focus on practical deployment, he participated in the development of Next-ViT: Next Generation Vision Transformer for Efficient Deployment in Realistic Industrial Scenarios. This hybrid CNN-Transformer architecture combined deployment-friendly blocks with a novel stacking strategy to optimize the latency-accuracy trade-off, contributing to the field of Hardware Aware Design for On Device Deep Learning.
By 2024, Wang's focus expanded to the acceleration and optimization of generative models, particularly diffusion models. He collaborated on Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesis, a distillation framework designed to enable high-quality image generation in very few steps. By combining ordinary differential equation (ODE) trajectory preservation with human feedback and score distillation, this work advanced the research direction of Accelerating Diffusion Sampling via Distillation and Trajectory Rectification and touched upon the area of Cross-Architecture Knowledge Distillation for Efficient Reasoning.
In 2025, Wang continued to investigate efficient generative modeling. He participated in the development of VarFlow: Proper Scoring-Rule Diffusion Distillation via Energy Matching, which proposed transforming multi-step diffusion models into one-step generators by minimizing a sample-based energy distance rather than relying on unstable score-matching objectives. That same year, he contributed to LABridge: Text–Image Latent Alignment Framework via Mean-Conditioned OU Process. This research introduced a structured, text-aligned prior using a mean-conditioned Ornstein–Uhlenbeck process to pull the generation process toward desired text semantics, improving both sampling speed and text-to-image alignment. Additionally, he explored alternative architectures for low-level vision tasks, participating in the design of VmambaIR: Visual State Space Model for Image Restoration, which applied visual state space models to image restoration problems.
Most recently, in 2026, Wang's work has addressed the computational bottlenecks of deploying iterative generative models on edge devices. He collaborated on Error Propagation Mechanisms and Compensation Strategies for Quantized Diffusion Models, investigating the hidden flaws of post-training quantization in diffusion models, specifically how errors accumulate during the iterative denoising process. He also contributed to Motion-Aware Caching for Efficient Autoregressive Video Generation, a framework that accelerates video synthesis by skipping redundant computations on a per-token basis using frame differences as a proxy for motion.
Current Directions
The most recent papers in the evidence indicate a strong focus on the efficient deployment and acceleration of large-scale generative models. Recent work investigates strategies to drastically reduce the computational burden of diffusion and autoregressive models without sacrificing output quality. This includes exploring advanced distillation techniques, such as energy matching, to achieve high-fidelity one-step generation, as well as motion-aware caching mechanisms to accelerate video synthesis by bypassing redundant token computations.
Additionally, Wang is currently exploring the intersection of model compression and generative stability. This involves analyzing and compensating for error propagation in quantized diffusion models, ensuring that post-training quantization remains viable for iterative denoising processes on edge devices. Alongside these acceleration efforts, recent research also examines alternative architectural paradigms, including the application of visual state space models to image restoration and the use of structured, mean-reverting stochastic processes to improve text-image latent alignment.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake