In our series titled Inside the Black Box, we aim to explain in simple terms the technical intricacies of the Alias platform. Our goal is to provide you, the artist, with a clear understanding of how our artificial intelligence (AI) systems operate behind the scenes.

Image created using Ornamental Ethnofuturism by Big Dan on Alias Studio

Inside the Black Box #1 : Understanding how captions impact Alias training
Inside the Black Box #2: Generating images on Alias without prompt engineering — or just a pinch!
Inside the Black Box #3: Providing the right input to generate the right Alias aesthetic

In this fourth installment, we’ll explain how training iterations work from both a theoretical and practical perspective so that you can make the most out of them. Unlike other platforms, Alias lets you understand what is happening by offering to visualize each iteration output during training. This, along with our testing studio, allows you to choose the most accurate result — not just the final version, but any along the way.

This piece will be illustrated with artworks by the late Julie de Graag, a Dutch artist renowned for her detailed woodcuts and engravings, often featuring natural themes. We will train an Alias for her work, carefully reviewing each iteration to select the most accurate rendition.

A 10-iteration training process

Do you remember the red token representing the style of the Alias from our last installment? Well, this token needs to be trained using the images from your dataset.

Training aims to mathematically capture the style of the images with minimal errors. On the Alias platform, this process unfolds over 10 steps, known as iterations. By the end of training, the AI model should be capable of matching the images as closely as possible to the style defined for your Alias.

To create this unique Alias token corresponding to your image dataset, we first use textual inversion. This technique enables the model to capture the style of the dataset images with a 768-dimensional vector. The vector starts with random values and is adjusted during the first 5 training iterations in order to find values with the lowest error rate possible. Think of it like this: the algorithm is mapping your artistic style, but instead of placing it on a simple 2D or 3D canvas, it’s positioning it within a vast, 768-dimensional space.

To allow the model to fully capture a style over 10 iterations, we utilize a technique called pivotal tuning. The latest version of the Alias token is used in conjunction with the images and captions to train the low-rank adaptation (LoRA) for the first 5 iterations. This allows to fine-tune the LoRA part while keeping it flexible. The first phase is like a global search on the map, finding where it best fits.

In the final 5 iterations, only the LoRA is updated, while the token’s value stays fixed, allowing the model to “fixate” on the style. This final part is akin to finding the exact location in the area that was decided during the first part.

Training is performed on every image in your dataset for each iteration.

The iterations in practice

Now that you have a better understanding of the theory, let’s examine training iterations of an Alias we’ve created using some of de Graag’s artworks.

In our training dataset, we uploaded 20 images by de Graag.

If you need a refresher on this step, take a look at our previous installment for best practices on training datasets.

Observing iterations

Once you have uploaded your Alias images in the dataset, you will be asked to select 12 prompts. Predefined prompts or custom prompts are available. For each of the 10 training iterations, 12 images will be generated based on those prompts.

The purpose of generating images based on those prompts during training is to help you understand the progress and learning process of the model. While these images aren’t directly part of the training itself, they serve as visual feedback for you to see how the model’s understanding evolves over time. You can observe how the model improves in interpreting prompts for each iteration.

The training can be launched once it has been completed. During each iteration, you will see the artwork generated for each testing prompt.

Iteration outputs can still be accessed after the training is completed by clicking on the iteration number. Hovering over each image shows the testing prompt.

Beyond the 12 prompts: Advanced testing of each iteration

Following the training, you can test the iterations to find the best one using the “Test now” button.

You can also find all the images for every iteration there.

To test each iteration individually, select it and enter prompts in the testing studio. You can fully leverage the Alias studio’s capabilities, including choosing the number of images you want it to generate.

Our test result is as follows:

Selecting the right iteration

Now you are able to choose between iterations. Keep in mind that the last one does not always represent the best option for you. With each iteration, the model becomes more and more fixated on the training dataset. Therefore, the 10th iteration with the fewest errors may not be the most flexible. Instead, the 8th or 9th iterations could better match your preferences.

That’s it! Your Alias is ready to be published!

Continuing training

Keep in mind that choosing an iteration from those first 10 is not your only option. And if you’re not completely satisfied with the results, you can always refine your model further.

Here are a few ways to do this:

Adjust the bucket influence in your training dataset
Implement prompt injections
Retrain the Alias from scratch

These methods can help you achieve the results you’re looking for and fine-tune the model to better meet your needs.

Iterating on the perfect Alias

Understanding the intricacies of training iterations on the Alias platform empowers you to take full control of the creative process. By breaking down each step, from the initial token training to the fine-tuning with LoRA, and providing you with the ability to visualize and test each iteration, we ensure that you can select the most accurate and suitable output for your artistic needs. Whether you’re aiming for precision or flexibility, you can make informed decisions at every stage.

As you continue to explore and experiment, you’ll find that the key to mastering the Alias platform lies in leveraging these iterations to refine your work, much like de Graag did with her intricate woodcuts.

Stay tuned for more insights from our Inside the Black Box series, where we will continue to explore and demystify the technical aspects of the Alias platform.

Alias Studio
Website

Follow Alien
Discord Community
Linkedin
Instagram

Medium Alias by Alien

Inside the Black Box #4: Getting the most out of training iterations was originally published in alien.papers on Medium, where people are continuing the conversation by highlighting and responding to this story.

Long form papers

Why open source is key to AI business

Navigating the uncharted waters of Generative AI in the creative industries.

Sign up for our newsletter