The Beauty of Systems— TEV1 Algorithm Breakdown

"DREAMLAND" - WINTER COLLECTION 2024

The TEV1 algorithm transforms regular images prompts into eye-bending visuals using generative AI and combinatorics.

Here’s how it works:

1. Basic Rules: A fixed number of artists, styles and features.

2. Prompt Engineering: We use prompt engineering techniques to transform these static features into unique visuals.

3. Combinatorics: The mathematical study of counting, combination, and permutation ensures that even slight changes in image features produces a vast array of different and unpredictable outputs.

Conway’s Game of Life

A classic example of a simple system that results in complex behaviours is Conway’s Game of Life. It is defined by a few straightforward rules:

1. Cell States: Each cell in a grid can either be alive or dead.

2. Neighbour Interactions: Each cell interacts with its eight neighbouring cells.

3. Rules:

• Any live cell with fewer than two live neighbours dies (underpopulation).

• Any live cell with two or three live neighbours lives on to the next generation.

• Any live cell with more than three live neighbours dies (overpopulation).

• Any dead cell with exactly three live neighbours becomes a live cell (reproduction).

Despite the simplicity of these rules, the patterns over time can lead to highly complex behaviours and structures, such as stable formations, oscillators (patterns that cycle indefinitely), and spaceships (patterns that move across the grid).

This understanding was fundamental in trying to create an algorithm based on simple rules to create unique digital art.

Infinite Combinations

In TEV1, we can model the diversity of artwork creation through combinatorics:

• 88 artists

• 84 art styles

• 25 compositions

• 52 orientations

We would like to add more artists and art styles but the artists included in the list are most likely to be known the image generation model. It is very unlikely inputting local or very boutique artists will produce anything coherent. The image model will interpret them as random names with no associated art style to reproduce. The same goes for compositions and orientations.

To calculate the total number of distinct artworks possible:

• Choosing 2 distinct artists from 88:
  C(88,2)=3,828C(88, 2) = 3,828C(88,2)=3,828 ways

• Choosing 1 orientation from 52:
  52 ways

• Choosing 1 composition from 25:
  25 ways

• Choosing 2 distinct art styles from 84:
  C(84,2)=3,486C(84, 2) = 3,486C(84,2)=3,486 ways

Multiplying these combinations together:

That’s over 17 billion unique artworks, demonstrating the mind-boggling power of combinatorics in generating visual variety.

Again, 25 compositions and only 88 artists does not sound like much variety can arise, but even a few rules such as taking fixed numbers from each set of features allows for a near infinite combination of images to be generated.

Endless Training Data and Examples

Over the years, we’ve gathered tens of thousands of viable image prompts to enhance our algorithm. These new image prompts continuously refine the algorithm through quality evaluation and parameter adjustments. Currently, we are developing a continuous delivery pipeline to integrate past descriptions into the live prompt generation process.

This will mean future versions of the TEV1 or just the TEV considering the TEV1 refers to the first version of the ‘Theme Explorer’ will be getting better in real-time as users interact with the tool.

Simple Rules = Complexity

As discussed above, simple rules leads to complexity.

"DREAMLAND"

Take this ‘DREAMLAND’ artwork for example, it appears like a simple ai-generated artwork but if you look closely, you can see there is slight spiral distortion in the left side of the canvas in the direction of the sun. This is in contrast to the right side where there is only a small amount if any distortion.

If we examine the image prompt, we can identify the reason for this visual effect.

“Fauvism, Digital Painting of Sunlit clouds with cotton candy hues swirling against an azure sky, creating a whimsical and dreamlike atmosphere. influenced by Henri Matisse, Wassily Kandinsky, Fluid, Tilted perspective, and Soft pinks, light blues, warm yellows, and pastel purples”

Words like ‘Tilted perspective’ and ‘swirling’ are the key components giving the artwork its tilted and distorted effect.

Creative Prompt Engineering

We employ a ReAct prompt framework to democratize the process of creating an artwork.

Each agent has their own role such as researching and exploring ideas while other agents will have the roles of prompt structuring and validation.

We have talked about this process more extensively in another article which you can access here.

In addition to ReAct, we utilize few-shot prompting and creative role assignment to optimize the performance of each agent.

Bringing it All Together

With TEV1, we present one of the most beautiful and sophisticated algorithms currently available, consistently producing high-quality art that draws inspiration from a comprehensive dataset and refined processes.

You can try it for free here.