Mitigate Data Replication in Diffusion Models

(05/2023-09/2023)

  • Developed dual fusion enhancement technique to mitigate data replication in diffusion models.
  • Constructed robust metrics to evaluate the trade-off between caption generality and data replication.
  • Reduced replication by 43.5% compared to the original model while maintaining the diversity and quality.
  • Submitted a paper to 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (Accepted) (EI).