DLight model still blurry after 400k iterations

[iguana314]

Hi,

The issue is pretty much as the title says. The picture is still completely burry. I ran the model for a week at we're at 395k iterations which seems enough. The mask type I'm using is VGG-obstructed.
I have about 1200 pictures from video A and B. Is it just a lack of data (though I thought the blurriness would go away after so many iterations...)

Any suggestions?

[torzdf]

*Generally* still blurry is down to data issues, or a requirement to train longer.

If the image is a close-up then there is likely to always be some blurring due to upscaling.

[sergor]

Hi,

Like you I trained still image-sets with the Dhl128 model. On 250K I made the first conversion - It was.. well not good!

I then did the following:

From training A -> B, I switched training B -> A. Focus on the two different sides has vastly improved conversion which was done at 350K and 450K iterations.

Matt (@torzdf) may agree here that if you pay attention to the numbers (example 0.11000, drops to 0.9000 then goes to 0.15000) it tells you that the model is still very much engaged in the training process and still "learning".

So my suggestion is when you have trained A->B 400K iterations, switch to B->A it also helps strengthen foundation of the model causing the results to become less corrupt in final images.

It also helps to tweak conversion settings (using gaussian, normalized etc.) to find the optimal result. But using still images is truly a long process and takes A LOT more training then if doing movie sequences.

I my experience (limited) that is.

Keep at it. Will share my results when reaching the 1.000.000 mark.

/sergor

[Helioses]

I then did the following:

From training A -> B, I switched training B -> A. Focus on the two different sides has vastly improved conversion which was done at 350K and 450K iterations.

So my suggestion is when you have trained A->B 400K iterations, switch to B->A it also helps strengthen foundation of the model causing the results to become less corrupt in final images.

/sergor

Is there a trick to do the A>B to B>A swap? Is there a checkbox anywhere, or do you change/swap the directories and alignment files in the training tab?

[bryanlyon]

Do not do that. It is the exact opposite of what you should do for a quality swap. (Keep A and B separate)

[dheinz70]

My experience with DLight gives good A quality, and totally washed out B quality. Dunno if that helps you. 192 SAE and Villain give me the best results.

[bryanlyon]

DLight is an "unbalanced" model. It's meant specifically for B's face swapped onto A. It will only work well in that direction.