LinCa: Accelerating Diffusion Models via Learnable Decomposed Feature Caching

JinShan Liu *, Haoran Qin *, Xiaobing Tu, Jiacheng Liu, Jiahui Hu, Zhengan Yan, Yukun Xie, Kerui Shen, Jinkui Ren, Yuqi Lin, Xiantao Zhang, Linfeng Zhang ^†

TL;DR: We propose LinCa, a learnable feature caching framework for diffusion model acceleration. By decomposing cached features via invertible networks and applying differentiated predictors per component, LinCa achieves near-lossless quality at high speedup ratios on FLUX, Qwen-Image, and HunyuanVideo

Kiroshi: An Agentic Perception System for High-Accuracy Image Parsing

Haipeng Zhou *, JinShan Liu *, He Zhang, Xuequan Lu, Jun Ma, Lei Zhu ^†

TL;DR: We propose Kiroshi, an agentic perception system for high-accuracy image matting and segmentation. By training an Action Model with iterative refinement and MLLM post-training via within-context preference pairs, Kiroshi achieves fully automatic, state-of-the-art fine-grained image parsing.

Hide and Find: A Distributed Adversarial Attack on Federated Graph Learning

JinShan Liu *, Ken Li *, Jiazhe Wei, Bin Shi ^†, Bo Dong

[PDF] [Github]

TL;DR: We propose FedShift, a two-stage distributed backdoor attack on Federated Graph Learning that hides a learnable shifter during training and finds adversarial perturbations post-training, achieving state-of-the-art attack success while evading defenses with 90%+ reduced cost.

Multi-modal Service Robot Based on ROS Architecture

We built a multi-modal fully automatic home service robot powered by YOLOv8, MediaPipe, and InsightFace over ROS communication, achieving precise vision capabilities (facial recognition, object detection, action recognition, ACC > 95%) and interaction abilities (voice interaction, autonomous navigation, object grasping, near-zero failure rate).