62 lines
3.6 KiB
Markdown
62 lines
3.6 KiB
Markdown
## About learning Textual Inversion
|
|
|
|
[Textual Inversion](https://textual-inversion.github.io/). I heavily referenced https://github.com/huggingface/diffusers/tree/main/examples/textual_inversion for the implementation.
|
|
|
|
The trained model can be used as is on the Web UI.
|
|
|
|
In addition, it is probably compatible with SD2.x, but it has not been tested at this time.
|
|
|
|
## Learning method
|
|
|
|
Use ``train_textual_inversion.py``.
|
|
|
|
Data preparation is exactly the same as ``train_network.py``, so please refer to [their document](./train_network_README-en.md).
|
|
|
|
## options
|
|
|
|
Below is an example command line (DreamBooth technique).
|
|
|
|
```
|
|
accelerate launch --num_cpu_threads_per_process 1 train_textual_inversion.py
|
|
--pretrained_model_name_or_path=..\models\model.ckpt
|
|
--train_data_dir=..\data\db\char1 --output_dir=..\ti_train1
|
|
--resolution=448,640 --train_batch_size=1 --learning_rate=1e-4
|
|
--max_train_steps=400 --use_8bit_adam --xformers --mixed_precision=fp16
|
|
--save_every_n_epochs=1 --save_model_as=safetensors --clip_skip=2 --seed=42 --color_aug
|
|
--token_string=mychar4 --init_word=cute --num_vectors_per_token=4
|
|
```
|
|
|
|
``--token_string`` specifies the token string for learning. __The learning prompt should contain this string (eg ``mychar4 1girl`` if token_string is mychar4)__. This string part of the prompt is replaced with a new token for Textual Inversion and learned.
|
|
|
|
``--debug_dataset`` will display the token id after substitution, so you can check if the token string after ``49408`` exists as shown below. I can confirm.
|
|
|
|
```
|
|
input ids: tensor([[49406, 49408, 49409, 49410, 49411, 49412, 49413, 49414, 49415, 49407,
|
|
49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407,
|
|
49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407,
|
|
49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407,
|
|
49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407,
|
|
49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407,
|
|
49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407,
|
|
49407, 49407, 49407, 49407, 49407, 49407, 49407]])
|
|
```
|
|
|
|
Words that the tokenizer already has (common words) cannot be used.
|
|
|
|
In ``--init_word``, specify the string of the copy source token when initializing embeddings. It seems to be a good idea to choose something that has a similar concept to what you want to learn. You cannot specify a character string that becomes two or more tokens.
|
|
|
|
``--num_vectors_per_token`` specifies how many tokens to use for this training. The higher the number, the more expressive it is, but it consumes more tokens. For example, if num_vectors_per_token=8, then the specified token string will consume 8 tokens (out of the 77 token limit for a typical prompt).
|
|
|
|
|
|
In addition, the following options can be specified.
|
|
|
|
* --weights
|
|
* Load learned embeddings before learning and learn additionally from there.
|
|
* --use_object_template
|
|
* Learn with default object template strings (such as ``a photo of a {}``) instead of captions. It will be the same as the official implementation. Captions are ignored.
|
|
* --use_style_template
|
|
* Learn with default style template strings instead of captions (such as ``a painting in the style of {}``). It will be the same as the official implementation. Captions are ignored.
|
|
|
|
## Generate with the image generation script in this repository
|
|
|
|
In gen_img_diffusers.py, specify the learned embeddings file with the ``--textual_inversion_embeddings`` option. Using the filename (without the extension) of the embeddings file at the prompt will apply the embeddings. |