For our Christmas special this year, we thought it would be fun to take a look at the rise of AI in the creative world, specifically AI generated artwork. We will not be postulating about how much “intelligence” is at work however will say the images have exceeded the artistic ability on offer in the Grainstone Lee office. We chose a universe of 90 hedge funds and prop trading firms and generated a collection art inspired by the Chinese zodiac animal of our samples foundation year together with their headquarter address.
How the AI works
Text-to-image models are machine-learning models where the input is a natural language description used to generate an image to match the user description.
A user might offer a description as simple as "a cat" or "a cat drinking water", or it could even be something more elaborate such as "a black cat wearing a NASA spacesuit, floating in space. The Earth is in the background"
The AI behind these models combines a language model that transforms input text into a latent representation and a generative image model that generates the image.
The most advanced models in production such as OpenAI's DALL·E 2, Google Brain's Imagen and StabilityAI's Stable Diffusion have been trained on vast volumes of image and text data scraped from the web.
As the complexity of the prompts increases, the models either omit elements or incorporate new ones that were not included in the prompt, which could be owing to a lack of explicit training material, limited data representation, or a lack of 3D awareness.
For this blog, we have used Dream by Wombo.ai a Canadian AI company founded by Nir Kabessa. It uses two open-source neural architectures: VQGAN (Vector Quantized Generative Adversarial Network) and CLIP (Contrastive Language–Image Pre-training).
VQGAN + CLIP is a neural network architecture that expands on OpenAI's groundbreaking CLIP architecture published in early 2021. VQGAN is a deep learning model that blends convolutional neural networks (traditionally used for images) with Transformers (traditionally used for language) to generate images based on resemblance to other images following training on pure image data. CLIP is used to evaluate how effectively user-fed captions relate to images, following training on image-caption pairings from the web.
Katherine Crowson an artist and mathematician produced the Google Colab Notebook that integrated VQGAN + CLIP, which was influenced by Ryan Murdoch's BigGAN + CLIP.
The AI models work hand in hand: CLIP (the preceptor) gives VQGAN input on how to match the image to the text prompt, while VQGAN (the generator) generates the image and sends it back to CLIP. This procedure is repeated until the desired image is produced.
Finding the right prompt
After experimenting with company values, founders names and a variety of other generic values we settled on blending Chinese zodiac animals (based on our samples foundation year) together with the headquarter address.
For example, a prompt for Balyasny Asset Management, founded in 2001 and headquartered on Lake Street in Chicago, may be "Show an image of a snake in Lake, Chicago," omitting the word street in Lake Street to generate more fascinating images.
We have included a quick reminder of the Chinese zodiac animals, their years and the traits associated with each of them. We will leave it up to you to decide if the firms follow their traits.
Below is the sample we have used in this blog post. An extended gallery of over 90 quant funds and prop firms can be found here:
1. Millennium Management
2. Citadel Asset Management
3. Susquehanna International Group
4. DE Shaw
5. Two Sigma
6. Jane Street
7. Point72 Asset Management
9. Citadel Securities
10. Man Investments
11. Balyasny Asset Management
13. Jump Trading
15. Virtu Financial
17. Tower Research Capital
19. Hudson River Trading
20. Squarepoint Capital
We couldn't pick a winner from the images however it did conjure up ideas for a top trumps style set of playing cards.... You can down the full image gallery using the link below.
Happy holidays from the Grainstone Lee team.
The full image gallery
If you would like to download a PDF of the extended gallery of over 90 quant funds and prop firms, please click here.