-
SparkDiffusion: Mitigating the High-Sparsity Trap
A unified framework for up to 265x single-GPU acceleration of visual generation.
-
CrossDistill: Balancing Quality and Diversity
Trajectory-level hybrid few-step distillation for high-quality and diverse generation.
-
RoLA: Rotary-Positioned Low-Rank Linear Attention
A rotary-positioned low-rank linear attention branch for efficient Diffusion Transformers.
-
RoPeSLR: Sparse-Low-Rank Attention for Diffusion Transformers
3D RoPE-driven sparse-low-rank attention for efficient Diffusion Transformers.
-
CR-Net: Cross-Layer Low-Rank Structure
Scaling parameter-efficient LLM training through cross-layer low-rank structure.
-
Mixture of Distributions Matters
Dynamic sparse attention for efficient video Diffusion Transformers.
-
MISA: Memory-Efficient LLM Optimization
Module-wise importance sampling for memory-efficient LLM pre-training and fine-tuning.