Features

This is a light-weight utility program for safetensors files written in Python only (no major external dependencies). Currently it can do the following:

Usage: safetensors_util.py [OPTIONS] COMMAND [ARGS]...

Options:
  --version    Show the version and exit.
  -q, --quiet  quiet mode, don't print informational stuff
  --help       Show this message and exit.

Commands:
  checklora    see if input file is a SD 1.x LoRA file
  extractdata  extract one tensor and save to file
  extracthdr   extract file header and save to output file
  header       print file header
  listkeys     print header key names (except __metadata__) as a Python list
  metadata     print only __metadata__ in file header
  writemd      read __metadata__ from json and write to safetensors file

The most useful thing is probably the read and write metadata commands. To read metadata:

    python safetensors_util.py metadata input_file.safetensors -pm

Many safetensors files, especially LoRA files, have a __metadata__ field in the file header that records training information, such as learning rates, number of epochs, number of images used, etc. You can see how your favorite file was trained and perhaps use some of the training parameters for your own model in the future.

The optional -pm flag is meant to make output more readable. Because safetensors files only allow string-to-string dictionary in metadata, non-string values must be quoted. Basically the -pm flag tries to turn this:

    "ss_dataset_dirs":"{\"abc\": {\"n_repeats\": 2, \"img_count\": 60}}",

into this:

    "ss_dataset_dirs":{
     "abc":{
      "n_repeats":2,
      "img_count":60
     }
    },

You can create a JSON file containing a __metadata__ entry:

{
     "__metadata__":{
          "Description": "Stable Diffusion 1.5 LoRA trained on cat pictures",
          "Trigger Words":["cat from hell","killer kitten"],
          "Base Model": "Stable Diffusion 1.5",
          "Training Info": {
                "trainer": "modified Kohya SS",
                "resolution":[512,512],
                "lr":1e-6,
                "text_lr":1e-6,
                "schedule": "linear",
                "text_scheduler": "linear",
                "clip_skip": 0,
                "regularization_images": "none"
          },
          "ss_network_alpha":16,
          "ss_network_dim":16
     }
}

and write it to a safetensors file header using the writemd command:

    python safetensors_util.py writemd input.safetensors input.json output.safetensors