research SparkDiffusion Unified framework for up to 265x single-GPU acceleration of visual generation. CrossDistill Trajectory-level hybrid few-step distillation for quality and diversity. RoLA Rotary-positioned low-rank linear attention for efficient Diffusion Transformers. RoPeSLR 3D RoPE-driven sparse-low-rank attention for efficient Diffusion Transformers. Mixture of Distributions Matters Dynamic sparse attention for efficient video Diffusion Transformers. CR-Net Cross-layer low-rank structure for parameter-efficient LLM training. MISA Memory-efficient LLM optimization with module-wise importance sampling.