Mr DeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
DSSIM in Faceswap
#1
I don't know about DFL, which I don't use (except that some of its models were ported to Faceswap in the last month), but I have heard on here that DSSIM is only useful for the first 15k iterations in Faceswap. Can someone explain what DSSIM does, and why this is so?

Does this mean that when creating a new model, I should routinely leave DSSIM on until 15k and then turn it off? Some of the Faceswap settings can't be revised after the model has been created, but I know that Warp landmarks can, and I presume that DSSIM can too.

Any further info on DSSIM gratefully received. As usual, up-to-date (or in-depth) guides are a bit scarce. Google hates searching for 'Faceswap', and most of the related technical terms, and generally shoots me straight to the academic papers where these features were first proposed; so it's hard to be very self-reliant when trying to learn this stuff!
#2
DSSIM is basically "Structural Similarity" which generally outperforms traditional loss functions. There is no need to switch it off and on as far as I know. However, if you wanted to change the loss mid-train, you can manually edit the state.json file in the model folder to switch it (I'd only change the loss in this file, anything else is likely to break the model).

The intro and conclusion in the wikipedia page explains in a fairly straightforward manner: You are not allowed to view links. Register or Login to view.

Forum Jump:

Users browsing this thread: 1 Guest(s)