Research on Personalized Image Diffusion Models with Adaptive Weight Selection (Challenge Cup National Third Prize)

• Extended the CVPR 2025 K-LoRA framework for training-free object-style fusion in diffusion models.
• Designed and analyzed K-LoRA++ scale-factor schedules, covering both linear and nonlinear variants.
• Ran cross-domain fusion and ablation experiments, showing that linear schedules provide the best balance of controllability, smooth transitions, and content fidelity.
• Supported the project’s practical implementation and downstream deployment.

SparseTSF-FFT: A Frequency-Domain Optimized Framework for Time Series Forecasting

• Led the PyTorch-to-MindSpore migration of SparseTSF, rebuilding the training and evaluation pipeline, implementing missing operator equivalents, and open-sourcing the code.
• Proposed SparseTSF-FFT, an FFT-based learnable filtering framework for long-term time series forecasting that replaces local convolutions while preserving the model’s sub-1k parameter budget.
• Reproduced weight-pattern visualizations, compared runtime and MSE across frameworks, and achieved stronger long-horizon forecasting results on ETTh1 and ETTh2.

GPU Optimization for All-Pairs Shortest Path (Solo Participant, National Champion in the CCF-TCARCH Computer Architecture Challenge)

• Designed a 3-stage blocked Floyd-Warshall architecture that heavily leverages shared memory to reduce global memory bandwidth bottlenecks.
• Applied kernel-level optimizations including fusion, dual-stream concurrency, double buffering, grid compression, and sync-barrier elimination, while cutting H2D transfer from O(V²) to O(E) through fast I/O and device-side reconstruction.
• Achieved a 52%+ end-to-end speedup, reducing total runtime from 11.50s to 5.42s through rocprof-driven profiling and hyperparameter tuning.