Welcome to your ultimate guide on how to effectively train your models using Extended Dreambooth. Whether you are working on a local PC, utilizing Google Colab, or accessing cloud computing platforms like Vast.ai, this step-by-step guide will light your path through the process. If your code skills feel a bit rusty, fear not! We’ll help you navigate through with ease, and we’ll also troubleshoot common issues along the way.

Table of Contents

• Notes by Joe Penna
• Setup
• Easy RunPod Instructions
• Vast.AI Setup
• Run Locally
• Configuration File and Command Line Reference
• Captions and Multiple Subject/Concept Support
• Textual Inversion vs. Dreambooth
• Using the Generated Model
• Debugging Your Results
• Hugging Face Diffusers

Notes by Joe Penna

Hello! I’m Joe Penna, known from my YouTube channel as @MysteryGuitarMan. As a feature film director, I created modifications to existing models to tailor them to my specific needs for training actors and props. This guide presents a community-driven resource for those interested in utilizing Dreambooth.

Setup

Your journey begins with setting things up, which involves selecting your platform of choice and preparing the framework.

Easy RunPod Instructions

To run your model on RunPod, follow these steps:

1. Sign up for RunPod and pick a secure or community cloud.
2. Choose an instance with at least 24GB of VRAM, like RTX 3090.
3. Edit Pod to update the Docker Image Name following the recent changes.

Vast.AI Setup

For those opting for Vast.AI, here’s how to get started:

1. Sign up for Vast.AI and add funds.
2. Select a valid Docker image, ensuring sufficient disk space.
3. Clone the repository using Git and follow along with instructions in the notebook.

Running Locally

If you prefer running your model on your local machine, you can choose between using Virtual Environments or Conda. Here’s how:

Setup – Virtual Environment

1. Clone the repository using Git.
2. Create and activate a virtual environment.
3. Install the necessary dependencies and start training using your command line.

Setup – Conda

1. Open Anaconda Prompt and clone the repo.
2. Install the environment and activate it.
3. Run the training command thereafter.

Configuration File and Command Line Reference

Sample configuration helps in training the model efficiently. You can refer to parameters such as class_word, max_training_steps, and others in your config file to achieve precise results.

Captions and Multiple Subject/Concept Support

Captions are a useful way to guide the training data. For instance, if you have an image titled img-001@a photo of token, you can customize it further using additional symbols to denote classes dynamically.

Textual Inversion vs. Dreambooth

This repository takes certain ideas from both Textual Inversion and Dreambooth, aiming to provide a unified solution with improvements.

Using the Generated Model

By training a model, you can generate diverse outcomes using similar prompts. For example, you might get various versions of an individual from different angles or styles with a proper training set.

Debugging Your Results

Here are common hurdles you may encounter during the training:

They don’t look like you at all!

Ensure your prompts include both the token and class; for instance, JoePenna person instead of just JoePenna. You may also need to train longer or enhance your training dataset.

They sorta look like you, but exactly like your training images

In this case, you may have trained for too long or used overlapping images. Try adjusting your prompt to focus on clarity in what you want to generate.

They look like you, but not when you try different styles

If you want the generated images to appear in a variety of styles, you might need a longer training time or a broader set of training images.

Troubleshooting

Common solutions include checking the configuration files and ensuring your training images are diverse and varied. Avoid using a single token for image generation, as context helps immensely.

For more insights, updates, or to collaborate on AI development projects, stay connected with fxis.ai.

At fxis.ai, we believe that such advancements are crucial for the future of AI, as they enable more comprehensive and effective solutions. Our team is continually exploring new methodologies to push the envelope in artificial intelligence, ensuring that our clients benefit from the latest technological innovations.

Hugging Face Diffusers

If you are looking for an alternate avenue, you can explore how Dreambooth has been integrated into Hugging Face Diffusers for enhanced training possibilities.

Good luck on your journey with Extended Dreambooth!