There’s no need to get my training data, simply start up character creator with a headshot plug-in and run a bunch of AI-generated front facing portray face images through it and you have a good training data set.

@Polygon At first, start with a small dataset of rather similar images with subtle changes. Let’s say something like 10 images for each main keyword. Like woman, man, old, young, make-up, european, east Asian … beard and so on. Make sure that you are tagging the images with the triggering keyword in the text file as the first word inside your description text file that has the same name as the image but is a .txt, after that in the text file separated by commas, the main attributes that you like to have associated with the image so the previously mentioned woman, man, old, young, make up and so on. Make sure that all your initial training data has the same background colour for the area that is outside of the UV boundaries. In order to train the model to not draw details there and get confused. definitely remove any kind of noise, watermarks, or whatever might be outside of the used UV space on your texture model page. If you don’t do this, your training will not generate a clean model that you can properly work with and expand upon later.

do the training on this dataset, run it with a rather low number of iterations, and often generate sample images and intermediate checkpoints so that you can see the point where the iterations over train the network and the results start to degenerate. It’s really easy to unintentionally overtrain a network that is dealing with such delicate changes like those in diffuse textures for a 3D model.

After you’ve trained the first checkpoint and found the sweet spot where it’s not over not undertrained, you have a foundation model that you can use for later reinforcement. Keep all drastic changes to the facial features of your training data set for the reinforcement face when you train on top of an already trained model. Because if your source images in the first stage already have a strong variation, you’re reducing the later ability to edit and utilise the retraining capability.

Don’t be afraid to cancel a training if you see it going off the rails. it’s better to start 20 times over until you find the right configuration than to waste hours on calculating an already failing model training to the bitter end, where it becomes artefacts and noise. If you see signs of overtraining or degeneration, they are just what I did say, and there’s no point in continuing training.

Doing something like that for DAZ is interesting however they’ve changed the UV layouts between generations so there is a difference between the V4 and Genesis characters and so on so since they don’t have the same UVW layout you would have to pick and choose one of them to do the training on since mixing different layouts in the same training will screw up the alignment.

@bee @Clare3Dx At first, start with a small dataset of rather similar images with subtle changes. Let’s say something like 10 images for each main keyword. Like woman, man, old, young, make-up, european, east Asian … beard and so on. Make sure that you are tagging the images with the triggering keyword in the text file as the first word inside your description text file that has the same name as the image but is a .txt, after that in the text file separated by commas, the main attributes that you like to have associated with the image so the previously mentioned woman, man, old, young, make up and so on. Make sure that all your initial training data has the same background colour for the area that is outside of the UV boundaries. In order to train the model to not draw details there and get confused. definitely remove any kind of noise, watermarks, or whatever might be outside of the used UV space on your texture model page. If you don’t do this, your training will not generate a clean model that you can properly work with and expand upon later.

do the training on this dataset, run it with a rather low number of iterations, and often generate sample images and intermediate checkpoints so that you can see the point where the iterations over train the network and the results start to degenerate. It’s really easy to unintentionally overtrain a network that is dealing with such delicate changes like those in diffuse textures for a 3D model.

After you’ve trained the first checkpoint and found the sweet spot where it’s not over not undertrained, you have a foundation model that you can use for later reinforcement. Keep all drastic changes to the facial features of your training data set for the reinforcement face when you train on top of an already trained model. Because if your source images in the first stage already have a strong variation, you’re reducing the later ability to edit and utilise the retraining capability.

Don’t be afraid to cancel a training if you see it going off the rails. it’s better to start 20 times over until you find the right configuration than to waste hours on calculating an already failing model training to the bitter end, where it becomes artefacts and noise. If you see signs of overtraining or degeneration, they are just what I did say, and there’s no point in continuing training.

Character studio has a feature that can match the body texture to the face texture if you have the skin generation things available.

Thank you for the link I have a look at that git repository

Image to image with depth and controlnet with depth or normal extraction work with any checkpoint. However, the fundamental problem is that as soon as you divert too far or use a complex clip text instruction, the texture pages degenerate. That’s why training a checkpoint that already knows where to place what main feature of the texture page gives better results and can of course, simply can be combined with the aforementioned depth map and as well as, the features of controlnet.

see https://civitai.com/models/18777?commentId=57241&modal=commentThread

Use your layout reference base face in image 2 image?