Faceswap Forum

martinf

I am wondering if it is possible top start training with Penalized Mask Loss ON then, after facial features are doing well, turn it OFF in order to improve some of the non-trained masked areas?

martinf

I want to do a swap in which the B face wears glasses and they are very much a part of the person's identity but the A face does not wear glasses.

Can this be done without going through crazy hand drawn mayhem?

Thanks,

martinf

I am missing a lot of faces that are frontal but tilted quite a bit. I have tried 0 (thinking it might defeat the filter) and all the way up to 90 and I see no difference in the output at all. I have also typed in a higher value and get the exact same results. Maybe I am misunderstanding what this f...

martinf

It appears that the 'roll' filter tosses out some faces that are very tilted, but that are still fairly frontal views. Might be wrong on that. It would be NICE to have a function that instructs the extractor to find a face - period. Like saying "Hey... I can see the face clear as day... at leas...

martinf

Does the alignment file simply tell the model where to find the face, then the model analyzes the pixels for the swap... or are the alignments playing a larger role in the swap process? Ie., Does the model look to the alignments for the face generation or only to the underlying video... after the al...

martinf

Cannot get SYM384 to behave. Using the stock preset. Everything else standard. 3080Ti on a strong machine with tons of RAM. I either end up with bright solid yellow and magenta blocks or solid white blocks. Have no clue where to start to adjust this to make it work.

martinf

I think I read something here somewhere about this but I cannot find the right search terms to dig it up so here is my likely redundant question... Is there any value in training without a mask? It would seem this might enable swaps with more difficult to fit face types especially in profiles. I kno...

martinf

rolnor wrote: ↑Wed Aug 31, 2022 1:52 pm
I tried Phaze-A with the 1024 presets and 2048 output res. its running fine with my 1080Ti at batchsize 70. It seems fantastic! I can already see start of a good model at 1750iter.

Impossible.

martinf

It got to where anything I tried to run from Phaze-A was corrupted. Stojo gave me nothing but bright squares then a NAN after about 1500 iterations at batch size of 3. Then... other models also started to crap out. Powered down and turned power supply off for 30 seconds and restarted. Other models w...

martinf

Not just OOM, but a laundry list of errors involving the keras decoder. What I did learn is that starting training back up and letting it run for a few iterations clears whatever was screwing up the preview function. Prior to doing that the preview would hang no matter what project/model I would fee...

martinf

The 256 Stojo Dyn model is crashing here on a batch size of 3, Windows 10, 3080Ti.

And, after it crashes, the preview function of Faceswap stops working altogether until I reinstall. Not sure if this is an edge case or what.

martinf

I can run that model with batch size of 3 without issue. If I go to 4 it will run for a while then any stress at all on VRAM creates an OOM. This is on a Win 10 Pro box with 3080Ti and 80 gigs RAM and very little else running. Again... SYM384 Phaze_A with what I believe are default model settings bu...

martinf

Running the SYM384 model and I am running out of VRAM at batch size of 4 with a 3080Ti.

I am wondering if adding a second GPU only to be used to drive my display might free up some VRAM on the 3080Ti?

Anyone know if this is worth exploring?

martinf

I just had a NAN with NAN protection on and the model was still toast for some reason. I turned mixed precision off, adjusted the epsilon and tried to resume. The next preview update was garbage...

Is NAN protection supposed to save the model or just save electricity by shutting down?

martinf

Thanks for the clarification. It seems as if my issue was mostly dealing with image sets that had a plethora of straight on to 3/4 angles in yaw but a much lower percentage of profiles. The profiles were there, but because they were about 5% of the totals they seemed to be lagging badly. Brings up a...

martinf

Is training a function of epochs or iterations? Ie., are these two statements correct? Given a small sample of A and B images (say, 100), training will be fast... but likely poor. Given a large sample of A and B images (say, 5000), training will be slow... but likely of much higher quality. I ask be...

martinf

Using the Dfl Sae model at the moment. The graph shows numbers that are about 33% of the numbers posted at the save intervals. Is this something peculiar to this model as there are three graphs (32, 64 and 128). Are those to be added together to arrive at the save interval number?

martinf

I tried the setup that had the fully shared FC and the identity bleed was ridiculous. The progression from A to B after 60K iterations looked like I had simply applied a gaussian blur to the original image. There was no indication that a swap was even taking place. This is frustrating to just have t...

martinf

Should I just use the included presets or is there something that has been discovered that creates a better launching point?

martinf

I would agree with you, but any occasion to expand the mask a tad and it is a very noticeable thing. I'll mess with it a bit. For sure this ends up in a trained model in which the lpips_alex is added into the training midway.

Faceswap Forum

Search found 28 matches

Can Penalized Mask Loss be changed in the middle of training?

Is it possible to swap a B face wearing glasses with an A face that does NOT wear glasses?

Can the 'roll' filter be bypassed?

Re: Alignments for faces at angles

Question about the conversion process - alignments versus pixels - need to understand

SYM384 Model Preset yielding solid color blocks after a few thousand iterations

Question about training WITHOUT a mask

Re: Phaze-A, 2080Ti

Re: [Guide] Introducing - Phaze-A

Re: [Guide] Introducing - Phaze-A

Re: [Guide] Introducing - Phaze-A

Re: Will installing a dedicate GPU in Windows 10 for my display only free up VRAM from a second GPU?

Will installing a dedicate GPU in Windows 10 for my display only free up VRAM from a second GPU?

Re: [Discussion] How to fix Mixed Precision causing NaNs

Re: Let me see if I have this right regarding training time...

Let me see if I have this right regarding training time...

Why are the progressively posted loss numbers very different from the ones in the graph?

Re: [Guide] Introducing - Phaze-A

Suggested starting point with Phaze-A for a 384-512 training done with 3080Ti in Windows?

Re: [Discussion] Notes on Loss functions