Faceswap Forum

Original Model, with a batch size of 192??

Not sure that's gonna get you very good results. I suggest Villain or DFL-SAE. I get the best results from them.

Is there any benefit to stepping down the batch size as the model learns? My general recipe is run at max batch size for a few days, cut that batch size in 1/2 for a few more. Then remove warping and cut that batch size in 1/2 again and run for a day.

What do you find works best? Thanks in advance.

Make sure you using the proprietary Nvidia driver on Linux? A default install uses the open source display drivers. Click the "Additional Drivers" icon and check from there. Also, the 450 nvidia driver works better than the 460 one. Just a guess.

Two similarly capable cards should work. The rate determining step will be the slowest card. Dunno if the 30x cards are supported yet.

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.

I get similar results. See viewtopic.php?p=4146#p4146

https://www.newegg.com/msi-mpg-x570-gam ... 6813144262

Another question, When using both PCIe slots do all MBs switch them to 8x8 or do some keep both the slots at 16x?

When I save a new project it adds the extension twice.

"realfacetest.fsw .FSW" is what it named the file, I only entered "realfacetest" in the name window.

The two bugs I've seen:

Changing the smoothing from 0.9 causes the stats to crash

It shows more iterations than the session has done. Hope that helps.

The fact it drops down to 8x8 tells me it is probably mostly hardware. Just thought it was weird that running two cards is almost exactly half as productive.

2000 iterations of a batch of 2 each on single and distributed. I doubt it was ever enuf data to clog the pipes. My feeling is there is some hardware limitation, but i suspect there is something else going on too.