Khái niệm cốt lõi
Incorporating noise priors enhances semantic image synthesis quality.
Tóm tắt
Introduction:
Semantic image synthesis aims to generate high-quality images aligned with semantic maps for applications like autonomous driving and robotics.
Challenges with Current Techniques:
GAN-based methods have not reached desired quality levels for practical sensor simulation applications.
Proposed Solution:
Developed specific noise priors encompassing spatial, categorical, and joint prior for inference, named SCP-Diff.
Results:
SCP-Diff achieves exceptional results on Cityscapes and ADE20K datasets, setting new benchmarks in semantic image synthesis.
Experiments:
Evaluation on multiple datasets shows superior performance of SCP-Diff over ControlNet and other state-of-the-art methods.
User Study:
User study confirms the higher quality and fidelity of images generated by SCP-Diff compared to ControlNet.
Thống kê
ECGAN achieves 44.5 FID on Cityscapes, while SCP-Diff achieves 10.5 FID.
SCP-Diff yields an FID of 12.66 on ADE20K.