MC-Sparse: Deconstructing and Closing the Dense-Sparse Attention Gap in Diffusion Transformers
Jiarui Chen, Zeqiang Lai†, Jiangshan Wang, Ziheng Ouyang, Ye Huang, Xiangyu Yue, Cewu Lu*, Chunchao Guo*
Website Paper Code

We present MC-Sparse, a training-free, token-level sparse attention framework. We analyze three sources of the gap between dense and sparse attention and address each in turn, achieving 1.80× denoising speedup for video generation and 2.32× for 3D asset generation.

Position Forcing: Self-Conditioning 3D Generation
Ziheng Ouyang, Zeqiang Lai†, Jiarui Chen, Jiangshan Wang, Yuhao Wan, Jingbo Gong, Xiangyu Yue, Hengshuang Zhao, Qibin Hou‡, Chunchao Guo‡
Website Paper

We present Position Forcing, which recovers token positions from predicted clean geometry latents and feeds them back as coarse-to-fine spatial guidance during denoising, improving single-stage 3D generation without a separate position generation stage.

Does Native 3D Texture Generation Necessarily Require 3D Assets for Training?
Jiangshan Wang, Zeqiang Lai†, Jiayi Guo, Xin Yang, Xin Huang, Jiarui Chen, Ziheng Ouyang, Chunchao Guo*, Xiangyu Yue*
Paper Project

We present Tex-Zero, a native 3D texture generation framework trained solely on 2D images. By constructing 3D training samples from image patches, it generates detailed 3D textures without using real 3D assets for training.

Elastic Diffusion Transformer
Jiangshan Wang, Zeqiang Lai†, Jiarui Chen, Jiayi Guo, Hang Guo, Xiu Li, Xiangyu Yue*, Chunchao Guo*
Paper Code

We present Elastic Diffusion Transformer (E-DiT), which achieves ~2× acceleration by dynamically allocating computation (depth and width) based on sample-dependent sparsity.

MEGS²: Memory-Efficient Gaussian Splatting via Spherical Gaussians and Unified Pruning
Jiarui Chen*, Yikeng Chen*, Yingshuang Zou, Ye Huang, Peng Wang, Yuan Liu†, Yujing Sun†, Wenping Wang
Website Paper Code

We present MEGS², a novel 3DGS compression framework that shifts the focus from storage reduction to the critical bottleneck of rendering memory, enabling real-time performance on edge devices and in web browsers.

Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation
Tianran Chen, Jiarui Chen, Baoquan Zhang*, Zhehao Yu, Shidong Chen, Rui Ye, Xutao Li, Yunming Ye
Paper

We present CDRA-SPT, a novel PEFT method that harnesses the complementary strengths of selection-based and reparameterization-based approaches.