The conversion effect between images is good during training, but it is much worse when predicting with the trained weights

Why are the images generated during training (in the index) converted well, but when the trained weights are used for inference, the conversion effect of these images is worse than that of the images converted during training? I can't figure it out.