For example, when training the class "man" you don't want the class "woman" to be affected as well. In Dreambooth training, reg images are used as an example of what the model already can generate in that class and prevent it from training any other classes. I don't know where this theory originates, but I find it to be misinformation. Something I've read on the internet, that I do NOT recommend, is using Stock Photos or otherwise Real Images as reg images. Save the images in the PNG format and put all of them in a "train" folder. In my testing, using only plain backgrounds gave a poorer result. Make sure to include images with a normal background (for example, of your subject in a scene). So for now I can't recommend using transparent backgrounds. Transparent backgrounds may also work, but can sometimes leave a fringe or border around the subject as can be seen here: ![]() For subjects, I've found that including samples with either black or white backgrounds helps immensely. Once you've collected photos for a dataset, crop and resize all the images into 512x512 squares and remove any watermarks, logos, people/limbs cut off by the edge of the picture, or anything else you don't want in your final model. In my trainings the sample or instance images consist of around 10-100 sample images from the show/movie/syle I want to train. Avoid fan art or anything with a different style, unless you're aiming for something like a style fusion. ![]() Ideally, only pick images from the show or artist you're training. When training for a specific style, pick samples with good consistency. Be sure to use high quality samples, as artifacts such as motion blur or low resolution will get picked up by the training and appear in the images you generate with your model. The Datasetĭataset creation is the most important part of getting good, consistent results from Dreambooth training. Make sure to download the Stable Diffusion 2.0 Base model and not the 768-v or any other model.Īfter you installed the dependencies and loaded the correct model you should be able to train a model just like before. Pip install -upgrade git+ transformers accelerate scipy
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |