Faceswap Forum

I am seeking better settings to try for swapping photo model A with photo model B. I have over 10k high quality, good lighting images for both.

Currently I am unable to get any sharp or at least not quite blurry and passable image swaps using 128x128 inputs, dfaker model, batch size of 40, 4000 faces for each set - went through each and every image to ensure the wire frame face matches the photo and it positioned correctly.

Reasably good hardware with an Rtx2080 and 32GB RAM.

Even with 500k iterations, the preview is noticably blurry. The improvement from 100k to 500k iterations was small but present. Previous attempts ran more than 1 million iterations while worse blurry previews than 100k with realface (used the original model and 16 batch size, 64x64 input)

What I am wondering is, what is the best training settings to try for large photos sets with good lighting? About how many iterations ballpark should I wait till I see reasonable sharpness or abandon to different settings? Is there anything else I can do to sharpen things up? Faceswap can do photos well and not just videos, right?

Yes Faceswap can handle images instead of videos.

Quality issues are almost always with data. Unfortunately it's impossible to guess what might be wrong without examining your data or results. If you can, please post a training preview so we can evaluate your data and results.

Here is a typical level of blurriness -- this is after 650k iterations at the settings described above.

And here is a typical level of quality of source material before extracting the faces

I am not particularly attached to the models -- just grabbed easy to find, abundant photos for source material for trying out face swapping. But the process doesn't seem to be working so if the source material is a problem...what parameters should i look seek as better material? Personally, i thought, good lighting, clear photos, decent resolution, someone made packs of zip files, done. Lets start extracting and training.

I also haven't seen any screenshots of a completed training model preview to see what the level of detail would look like.

Those previews should almost definitely look better than that at 650k on dfaker (assuming a reasonable batch size). As Bryan says, blurriness is almost almost down to data quality. Variety is key.

So a batch size of 40 is too low for dfaker? And for variety of images, is 8000 individual photos in both the swap and original sets, so 16000 photos total, you don't think there is enough variation in that?

I can try and add to these photo sets or create a separate group of data to train on. What is the process for changing the original and swap data, using the same people and continuing the training with what the model has learned already?