A diffusion trainer for latent + embed only datasets.
Currently supports SD3.

To create a latent + embed dataset from images and text files:

1. Run create_bucket_list.py dataset_path to create a bucket list of the images.
2. Run create_latents.py model_path dataset_path new_latent_dataset_path --model_type sd3 --image_ext .jpg to create latents from the images.
   You can delete the original images after this step if you want, we don't need them after creating latents.
3. Run create_embeds.py model_path dataset_path new_embed_dataset_path --model_type sd3 --text_ext .txt to create embeds from the texts.
   You can delete the original texts after this step if you want, we don't need them after creating embeds.
4. Configure config.json with your desired configuration.
5. Run train.py path_to_config to start the training.

Running train.py will create a dataset_index.json file.
If you do any change to the dataset or the batch size, then remove this file before running train.py.

Example dataset_paths config:
["path_to_the_latent_dataset", ["path_to_the_embed_dataset"], repeats]

You can use multiple embed datasets with a single latent dataset as long as the folder structure and the filenames are the same:
["path_to_the_latent_dataset", ["path_to_the_embed_dataset_one", "path_to_the_embed_dataset_two"], repeats]