Page 1 of 1

maximizing input quality

Posted: Thu Feb 13, 2020 10:53 pm
by korrupt78
Let's say you want to use a face from a film, with maximum quality.

The highest quality is whatever the scene was originally record on - either a film negative, or some kind of raw DV tape. From there, the post-production process will end up re-encoding (and compressing) that original asset multiple times before it gets burned to a Blu-ray and purchased by you (or your friendly neighborhood torrent uploader).

After that, the video will be re-encoded *again* in the process of ripping that Blu-ray to a video file (unless you're literally copying/downloading the ISO).

Finally, when you run Faceswap Extract, it will get re-encoded *again* in JPG, which is yet another lossy format.

In theory, every re-encode reduces quality.

---

Now, I've read that input quality is far from the most important factor in maximizing the quality of final faceswap output, but it can't hurt to do the best possible job, right?

So, might it make sense to use a BMP training set instead of JPG to remove one of those re-encodings, and maybe even extract those BMPs direct from a disc image to avoid another?

Re: maximizing input quality

Posted: Thu Feb 13, 2020 10:54 pm
by bryanlyon
We don't use JPG. We use PNG which is lossless.

Re: maximizing input quality

Posted: Thu Feb 13, 2020 10:58 pm
by korrupt78
bryanlyon wrote:
Thu Feb 13, 2020 10:54 pm
We don't use JPG. We use PNG which is lossless.
I somehow failed to notice it was PNG instead of JPG. I also didn't know PNG was lossless - I though it supported lossy compression but was better at it - and less legally unencumbered. Good to know, that eliminates any concern about loss during the extraction step.

Aside from that, I'm just going to try to use the most upstream original sources I can get to feed into my pipeline.