Faceswap Forum

When using Phaze-a with the DNY256 preset, I am getting a gray outline around the converted face, usually on the top and sides. The swap otherwise looks amazing. The issue does not appear when using the same A/B data with DFL-SAE.

I tried using Reflect Padding with Phaze-A/DNY256, but then model doesn't seem to train at all (preview shows still solid grey converted face after 20k iterations).

Any ideas about what's causing the grey edges, and how to prevent it using Phaze-a?

I'd need to see an image to understand what you mean. This doesn't sound like an issue I've seen before.

Here is a comparison between DNY256 (200k it) and DFL-SAE (30k it)

Ok, It's to do with hairlines and available data. Reasons and potential mitigations are covered here:
viewtopic.php?p=7684&hilit=hairline#p7684

However, I am interested to know whether you are using identical data-sets and settings for each of the 2 models, because it is interesting to me that this doesn't happen with one model, but does with the other, which would be worthy of some further investigation.

I'll check out that thread, thanks.

A and B data sets are identical.

Starting with default DFL-SAE and DNY256 presets, I changed a few global options, but for both models:

Train-Global-Network-Mixed Precision: True
Train-Global-Loss-Mask: Vgg-clear
Train-Model-Dfl Sae: (No changes)
Train-Model-Phaze A: (Loaded DNY256 preset)

Batch Sizes: DflSae: 16 DNY256: 20

Converting:
Mask Type: Vgg-clear

Is there any other information that may help?

I can also try using the same data with other models or Phaze A presets.

Thanks for this, it is useful information. Just to confirm, these are both from presets used in Phaze-A, yes? (I mean that you used Phaze-A DFL-SAE preset, rather than the standalone DFL-SAE model?)

No - It was the standalone model of DFL-SAE.

I will now run a test of Phaze A + DFL SAE Preset though with the same data and settings to compare the three.

I have confirmed Phaze A with DFL SAE preset (loaded from model_phaze_a_dfl-sae-df_preset.json) behaves the same as the standalone DFL SAE model; the edges look good.

Let me know if there are any other tests I can run to help.