Jianing Deng中

SAF3R: Dynamic Sparse Attention for Feed-Forward 3D Reconstruction Transformers

J. Deng, Y. Li, J. Wang, S. Wang, T. Chen, H. Yang, J. Hu

ECCV 2026 [Paper] [Code]

A training-free dynamic sparse attention framework that exploits head-level heterogeneity to accelerate feed-forward 3D reconstruction transformers while preserving camera pose and geometry quality.

GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs

J. Deng, S. Wang, D. Wang, Z. Liu, T. Chen, H. Yang, J. Hu

ICML 2026 [Paper] [Code]

A mixed-precision quantization framework for MoE LLMs that jointly optimizes expert bit-width allocation and router adaptation, enabling ultra-low-bit quantization with minimal accuracy loss.

Quantization-Robust Unlearning through the Lens of Retain-Forget Loss Landscapes Interaction

J. Wang, J. Deng, S. Luo, Y. Li, D. Wang, J. Hu, H. Yang, S. Wang, T. Chen

COLM 2026

A quantization-robust LLM unlearning framework that flattens the forget-loss landscape and updates only forget-critical layers, preserving unlearning effectiveness while maintaining model utility.

Dynamic Mixed-Precision Routing for Efficient Multi-step LLM Interaction

Y. Li, J. Deng, J. Hu, T. Chen, S. Wang, H. Yang

arXiv 2026 [Paper]

A dynamic mixed-precision routing framework for long-horizon LLM agents that switches between full- and low-precision models at each decision step to improve the accuracy–cost trade-off.

FIER: Fine-Grained and Efficient KV Cache Retrieval for Long-context LLM Inference

D. Wang, Z. Liu, S. Wang, Y. Ren, J. Deng, J. Hu, T. Chen, H. Yang

EMNLP Findings 2025 [Paper] [Code]

A KV cache retrieval method that uses 1-bit quantized keys for fine-grained token selection, accelerating long-context LLM inference while matching full-cache accuracy.

Enhancing Facial Expression Recognition in Head-Mounted Displays with Synthetic Data

J. Deng, Q. Zhou, J. Hu

arXiv 2025 [Paper]

A training data synthesis framework that transforms abundant frontal-view facial images into head-mounted-camera-view data, improving facial expression recognition in AR/VR scenarios.

STHVC: Spatial-Temporal Hybrid Video Compression for UAV-Assisted IoV Systems

L. Chen, J. Deng, X. Zeng, L. Liu, Y. Wu, J. Hu, Q. Sun, Z. Shi, C. Zhuo

IEEE TCSVT 2025 [Paper]

A hybrid video codec that combines a space-time super-resolution network with conventional video codecs to achieve faster encoding and better rate–distortion performance.

STDF: Spatio-Temporal Deformable Fusion for Video Quality Enhancement on Embedded Platforms

J. Deng, S. Dong, L. Chen, J. Hu, C. Zhuo

ACM TECS 2024 [Paper]

An efficient video enhancement and super-resolution system for embedded platforms that leverages fast spatio-temporal deformable fusion for alignment and aggregation of neighboring frames.

Spatio-Temporal Deformable Convolution for Compressed Video Quality Enhancement

J. Deng, L. Wang, S. Pu, C. Zhuo

AAAI 2020 Oral [Paper] [Code]

A compressed-video quality enhancement method that introduces spatio-temporal deformable convolution for robust single-pass alignment and context aggregation across multiple frames.

Energy-Efficient Real-Time UAV Object Detection on Embedded Platforms

J. Deng, Z. Shi, C. Zhuo

IEEE TCAD 2019 [Paper] [Code]

An energy-efficient real-time UAV object detection system on NVIDIA Jetson TX2, achieving 28 FPS at 3 FPS/W and ranking 2nd among 53 GPU-track teams in the 2018 DAC System Design Contest.

TecoGAN-PyTorch

A multi-GPU PyTorch training framework for GAN-based video super-resolution (★ 240+), with a TecoGAN reimplementation that achieves a smaller yet stronger model than the official release.

↳ Adopted as the codebase by EGVSR (★ 950+), an efficient video super-resolution method.