About Pixel-Level Distillation Loss

[zhongtao93]

Have you tried using gt['rgb'] instead of gt['img'] to distll the student network? Or the gt['rgb'] is useless.

MobileStyleGAN.pytorch/core/loss/distiller_loss.py

Line 35 in 2d18a80

_gt_rgb = F.interpolate(gt["img"], size=_pred_rgb.size(-1), mode='bilinear', align_corners=True)

[bes-dev]

@zhongtao93 so, gt["rgb"] contain partial sums of the gt["img"], as we don't use aggregation of intermediate predictions like in StyleGAN2, it isn't correct to use gt["rgb"] here

[zhongtao93]

Since I want use MobileStyleGAN to blend anime and real face models, like stylegan in toonify. But I found this feature become weakened, especially when I reduce the channels of model.
Feature:

1. style codes lying in lower layers control coarser attributes like facial shapes,
2. middle layer codes control more localized facial features,
3. high layer codes correspond to fine details such as reflectance and texture.

[bes-dev]

@zhongtao93 I didn't try toonify pipeline on top of MobileStyleGAN. But if you have some experimental results it will be great if you share it.