Tips for newbie, blurry

If training is failing to start, and you are not receiving an error message telling you what to do, tell us about it here


Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for reporting errors with the Training process. If you want to get tips, or better understand the Training process, then you should look in the Training Discussion forum.

Please mark any answers that fixed your problems so others can find the solutions.

Locked
User avatar
squibcake
Posts: 2
Joined: Thu May 04, 2023 5:13 am

Tips for newbie, blurry

Post by squibcake »

Hi,

I've just started this fun yesterday and I'm doing my reading slowly, but I would really appreciate some pointers in the meantime, so I don't waste precious gpu-time.

I started training using the Original model with default settings, the only thing I changed (I think? ;) ) is batch size from def. 16 to 64.

  1. Is the original model enough for decent results in 1080p video? I'm so newb I just wanna make sure I'm not missing something.

  2. After 80k iterations the resulting swap is pretty blurry, and not really accurate (identity-wise): Image. I think I've seen better results in some tuts, but on the other hand, maybe it's to be expected?

  3. After 60k iterations I have changed my batch size to 16 (should improve quality/detail?) but I noticed loss values spiking after doing this, so I recovered my model and kept training on 64 (poof, 1 hour gone). I asume the loss sipke is to be expected? Should I try and reduce the batch size again to get better quality?

  4. Unfortunately, the first 30k iterations were done with 2 or 3 trash pictures (other person's partial face, no face) in the source face set directory. Luckily I spotted one of them on preview, deleted them and after that continued training. Can you comment on this? Is continuing training in such case a mistake? I would assume it gets ironed out in the process, but how much damage does it cause?

So, while I'm trying to learn, should I keep on training this one?

► Show Spoiler
User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 128 times
Been thanked: 622 times

Re: Tips for newbie, blurry

Post by torzdf »

squibcake wrote: Thu May 04, 2023 5:29 am
  1. Is the original model enough for decent results in 1080p video? I'm so newb I just wanna make sure I'm not missing something.

No, it isn't, but it is a good place to start out learning the process as you will get results relatively quickly. Original model is a 64px output, which will always be low-res compared to a 1080px frame.

  1. After 80k iterations the resulting swap is pretty blurry, and not really accurate (identity-wise): Image. I think I've seen better results in some tuts, but on the other hand, maybe it's to be expected?

Identity issues are normally data related (not enough variety). However, some people just do not like to swap well. I don't know why this is, as there appears to be no rhyme or reason to it, but it definitely happens.

  1. After 60k iterations I have changed my batch size to 16 (should improve quality/detail?) but I noticed loss values spiking after doing this, so I recovered my model and kept training on 64 (poof, 1 hour gone). I asume the loss sipke is to be expected? Should I try and reduce the batch size again to get better quality?

Loss will always spike on a stop/restart. This is normal. All that matters is that loss is trending downwards. Don't worry about the raw numbers.

  1. Unfortunately, the first 30k iterations were done with 2 or 3 trash pictures (other person's partial face, no face) in the source face set directory. Luckily I spotted one of them on preview, deleted them and after that continued training. Can you comment on this? Is continuing training in such case a mistake? I would assume it gets ironed out in the process, but how much damage does it cause?

If bad data is a small subset of your full dataset, don;t worry about it too much. Just remove the bad data when you can. These 'outliers' will tend to be averaged out.

So, while I'm trying to learn, should I keep on training this one?

Accept that your first swap will be terrible, and treat it as a learning experience. Everyone's first swap is bad!

My word is final

User avatar
squibcake
Posts: 2
Joined: Thu May 04, 2023 5:13 am

Re: Tips for newbie, blurry

Post by squibcake »

Thank you, I really appreciate what you're doing here. Awesome stuff, really.

Locked