About training detials in Table 3 about lafite HOT 5 CLOSED

Hi, results from table 3 are obtained by training from scratch, there is no pre-training.

from lafite.

PS, actually pre-training + fine-tuning is not guaranteed to obtain better results in this model. I think this is dataset- or domain-dependent.
In one of my previous experiment (not shown in the paper):
Pre-training with FFHQ + fine-tuning on MM-CelebA-HQ leads to better performance than training from scratch on MM-CelebA-HQ.
Pre-training on Google CC then fine-tuning on MS-COCO leads to worse results than training from scratch on MS-COCO.

from lafite.

The semi-supervised fine-tuning results in Fig.7(a) show that the Lafite achieves around 8.0 of FID trained on full COCO training data. Do the standard text-to-image Lafite models employ the semi-supervised fine-tuning method?

from lafite.

Results from table 3 are trained with only ground-truth pairs.

from lafite.

I read other issues and find that the contrastive regularizer for D and G plays an important role in Lafite. My confusion is solved. Thank you very much for your explanation. It was great help.

from lafite.