In-depth analysis reveals Lipschitz singularity issues in long-tailed datasets, suggesting a time-step sharing approach improves model stability and image quality.
Key Points
TCDM significantly improves noise prediction accuracy and feature transfer stability for tail classes.
Experimental results demonstrate superior performance on long-tailed datasets like CIFAR-100LT and CIFAR-10LT.
TCDM reduces the Lipschitz constant near zero, enhancing numerical stability and image generation quality.
This approach addresses the biases in class training behavior due to imbalanced data distributions.