[Textual inversion](https://arxiv.org/abs/2208.01618) is a method to personalize text2image models like stable diffusion on your own images using just 3-5 examples.
The `textual_inversion.py` script shows how to implement the training procedure and adapt it for stable diffusion.
## Running on Colab
Colab for training
[](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb)
Colab for inference
[](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_conceptualizer_inference.ipynb)
## Running locally with PyTorch
### Installing the dependencies
Before running the scripts, make sure to install the library's training dependencies:
**Important**
To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
**___Note: Change the `resolution` to 768 if you are using the [stable-diffusion-2](https://huggingface.co/stabilityai/stable-diffusion-2) 768x768 model.___**
**___Note: Please follow the [README_sdxl.md](./README_sdxl.md) if you are using the [stable-diffusion-xl](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0).___**
A full training run takes ~1 hour on one V100 GPU.
**Note**: As described in [the official paper](https://arxiv.org/abs/2208.01618)
only one embedding vector is used for the placeholder token, *e.g.*`"<cat-toy>"`.
However, one can also add multiple embedding vectors for the placeholder token
to increase the number of fine-tuneable parameters. This can help the model to learn
more complex details. To use multiple embedding vectors, you should define `--num_vectors`
to a number larger than one, *e.g.*:
```bash
--num_vectors 5
```
The saved textual inversion vectors will then be larger in size compared to the default case.
### Inference
Once you have trained a model using above command, the inference can be done simply using the `StableDiffusionPipeline`. Make sure to include the `placeholder_token` in your prompt.
For faster training on TPUs and GPUs you can leverage the flax training example. Follow the instructions above to get the model and dataset before running the script.
Before running the scripts, make sure to install the library's training dependencies:
It should be at least 70% faster than the PyTorch script with the same configuration.
### Training with xformers:
You can enable memory efficient attention by [installing xFormers](https://github.com/facebookresearch/xformers#installing-xformers) and padding the `--enable_xformers_memory_efficient_attention` argument to the script. This is not available with the Flax/JAX implementation.
Creating a training image set is [described in a different document](https://huggingface.co/docs/datasets/image_process#image-datasets).
### Installing the dependencies
Before running the scripts, make sure to install the library's training dependencies:
**Important**
To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
To be able to use Weights and Biases (`wandb`) as a logger you need to install the library: `pip install wandb`.
### Using your own data
To use your own dataset, there are 2 ways:
- you can either provide your own folder as `--train_data_dir`
- or you can upload your dataset to the hub (possibly as a private repo, if you prefer so), and simply pass the `--dataset_name` argument.
Below, we explain both in more detail.
#### Provide the dataset as a folder
If you provide your own folders with images, the script expects the following directory structure:
```bash
data_dir/xxx.png
data_dir/xxy.png
data_dir/[...]/xxz.png
```
In other words, the script will take care of gathering all images inside the folder. You can then run the script like this:
```bash
accelerate launch train_unconditional.py \
--train_data_dir <path-to-train-directory> \
<other-arguments>
```
Internally, the script will use the [`ImageFolder`](https://huggingface.co/docs/datasets/v2.0.0/en/image_process#imagefolder) feature which will automatically turn the folders into 🤗 Dataset objects.
#### Upload your data to the hub, as a (possibly private) repo
It's very easy (and convenient) to upload your image dataset to the hub using the [`ImageFolder`](https://huggingface.co/docs/datasets/v2.0.0/en/image_process#imagefolder) feature available in 🤗 Datasets. Simply do the following:
Before running the scripts, make sure to install the library's training dependencies:
**Important**
To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date. To do this, execute the following steps in a new virtual environment:
And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
```bash
accelerate config
```
For this example we want to directly store the trained LoRA embeddings on the Hub, so we need to be logged in and add the `--push_to_hub` flag to the training script. To log in, run:
```bash
huggingface-cli login
```
## Prior training
You can fine-tune the Würstchen prior model with the `train_text_to_image_prior.py` script. Note that we currently support `--gradient_checkpointing` for prior model fine-tuning so you can use it for more GPU memory constrained setups.
Low-Rank Adaption of Large Language Models (or LoRA) was first introduced by Microsoft in [LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685) by *Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen*.
In a nutshell, LoRA allows adapting pretrained models by adding pairs of rank-decomposition matrices to existing weights and **only** training those newly added weights. This has a couple of advantages:
- Previous pretrained weights are kept frozen so that the model is not prone to [catastrophic forgetting](https://www.pnas.org/doi/10.1073/pnas.1611835114).
- Rank-decomposition matrices have significantly fewer parameters than original model, which means that trained LoRA weights are easily portable.
- LoRA attention layers allow to control to which extent the model is adapted toward new training images via a `scale` parameter.
### Prior Training
First, you need to set up your development environment as explained in the [installation](#Running-locally-with-PyTorch) section. Make sure to set the `DATASET_NAME` environment variable. Here, we will use the [Pokemon captions dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions).
