Picture-to-picture face swap: poor quality output despite high-resolution sources

[mammuthus]

I tried to use high-resolution PNG sources for face swap, with default roop settings and without post-processing enhancers.
Original PNGs are very good-quality and have no visible problems (like compression artefacts, noise, etc.)
However, the result quality is really poor (it seems like it is JPEG with 1 of 10 quality set), and the picture is full of artifacts:

Is there no way to achieve good quality without post-processing, even if I use high-quality sources for both input and target faces?
Is this a problem with the model or with my settings?

[PP0815II]

As far as I understand it, the original image is resized to 512x512px after importing. I don't know which method is used (bicubic, bilinear etc).
I crop and resize the original photos to 512x512 beforehand with an external program. So that I have more control at this point (sharpness etc).
Nevertheless, a 512x512 head will look bad on a target photo with say 3000x3000px. Especially without any enhancer. The sweet spot will probably be when the head on the target photo is also about 512x512px. At least most of the photos I resized beforehand (souce AND target) look pretty good. Some are frighteningly convincing.