NaturalInversion: Data-Free Image Synthesis Improving Real-World Consistency

We introduce NaturalInversion, a novel model inversion-based method to synthesize images that agrees well with the original data distribution without using real data. In we propose: (1) Feature Transfer Pyramid which uses enhanced image prior of by combining multi-scale feature maps extracted from pre-trained classifier, (2) one-to-one approach generative where only one batch are synthesized generator bring non-linearity optimization and ease overall optimizing process, (3) learnable Adaptive Channel Scaling parameters end-to-end trained scale output channel utilize further. With our classifiers on CIFAR-10/100 show more consistent than works visualization additional analysis. Furthermore, outperform various applications such as knowledge distillation pruning, demonstrating effectiveness proposed method.