LORA VS Reactor

[m10renderz]

which is better, quick and more efficient, My personal choice ReActor, brilliant work, brilliant find, LORA for styles I get but for Face Reactor takes the trophy, in 1 minute trained through 2 images and consistent models for instagram,

[Woisek]

in 1 minute trained through 2 images and consistent models for instagram,

What do you mean with "trained"? This is just a simple face swap. 🤔
And the results can get good, but it highly depends on both of the source images. More often it's just ... well, a bit recognizable.

[Woisek]

@m10renderz
Thanks, I'm aware about the face model creation. But it uses only 1 image and not "add multiple images and train the face model for future use,"
And tbh, I don't get it, what the purpose is of that. If I can only use 1 image to "train" a model file, why even bother if I have the image and can use that ... ? What do I miss here?

make sure its well lit face,

More important: What minimum size should this face image be?

[m10renderz]

so to get a consistent model you train a face, and everytime you generate a new image, comes out with that face, helps me get a same model over and over again, yes I need to make sure that everytime I select Reactor, need to turn of the codeformer (included in reactor) , as it might change the eye color/shape or lips shape,

minimum size can be 500px, but well lit face, no shades , does a good job,

again to train a model is to avoid, faceswap in imgtoimg or in extra,
faceswaps automatically right while its done generating, i have tried doing lora's and so many times lora's not only come out with so many problems but also so time consuming vs a 1 min in face tool was why I was so impressed

[Woisek]

Okay, but still, what's the difference between using an image and a "trained" model?
I could understand the model generation, if we could take multiple faces of the same person, and it generates a face model from it. But with only 1 image ... I don't get the difference. 🤷‍♂️

[Gourieff]

Face model - is very useful when you don't want to search every time on your drive for the image you need, you just select it from the list and do the swap. And it also makes the first swap faster (coz there's no need for the analyser module to analyse the image)
In future updates I'm going to add an option of building a face model from a batch of images
This is not a training as it is - this is the face embedding with different additional parameters (in short - a "spatial" model of a recognized face) that can be used for recognition, identification or reconstruction purposes and as a consequence for the faceswap

[Woisek]

@Gourieff
Thanks for clarification. As for the search ... that's basically a good idea, but on the other side, depending on the use case, it can get confusing at some point. Be it to forget what name the face has, you want to use or, what I experienced already, using multiple variants of the same face. In this case, it's practically impossible to remember what face model is what face. In this case, only an enlargeable icon besides the name in the list would help.

I really like the idea of using multiple images for an "enhanced" face model. Depending on the source images, this may help to successfully swap angled faces better. I really hope we can use this soon. 👍 Thanks for considering this.