Why we cannot use real data for calibrating quantization parameter instead?

[kotomiDu]

If the target for generating fake data is to make the distribution to be same with the one from real data. Why we cannot directly collect the quantization parameter from the real data?

[kotomiDu]

As the paper mentioned, PSAQ-ViT can achieve better performance than Standard, which requires the real data, on all the aforementioned models, indicating that the generated images are even more effective than the real ones for parameter calibration. The main reason is that the sample generation is based on the prior information in the self-attention module, i.e., facilitating the distinction between foreground and background in images, and then when these samples are utilized to calibrate the quantization parameters, they in turn reinforce the functionality of the self-attention module, thus acting as positive feedback that can reduce the activation outliers to some extent and therefore improve the tolerance to parameter clipping