EMLART-GP - Tutorial GECCO 2025

This is a repository that serves as support for the Tutorial at GECCO 2025: Evolutionary Art and Design in the Machine Learning Era

Installation

Start by cloning the repository and enter the repository folder by issuing the following terminal commands:

git clone https://github.com/jncor/emlart-gp-tutorial.git
cd emlart-gp-tutorial/

This setup assumes and recomends a conda version greater than 23.3.1. The following commands will install the required packages to execute the code:

OSX arm64

conda create --name emlart-tutorial --file env-osx-arm64.txt
conda activate emlart-tutorial
python -m pip install git+https://github.com/openai/CLIP.git

OS-agnostic

conda env create -f environment.yml
conda activate emlart-tutorial
python -m pip install git+https://github.com/openai/CLIP.git

You will also need the weights for the models used by the provided scripts. You can get them from the following link: models

Just unzip the file ensuring that a folder called "models" is inside the cloned repository.

models/
tensorgp/
stablediffusion_examples/
emlart_gp.py
image_evaluator.py
image_evaluator_batch.py
laion_aesthetics.py
...

Usage

To execute the emlart-gp approach use the following command on the terminal in the repository folder:

python emlart_gp.py <starting random seed number> <# of runs> <# of generations> <text prompt>

E.g.

python emlart_gp.py 10 1 30 "sunset, bright colors" 

To execute the image evaluator based on Laion aesthetics machine learning model and cosine similarity with the a provided prompt via OpenAI Clip you should issue the following command:

python image_evaluator.py <path to the image> <text prompt>

E.g.

python image_evaluator.py "image_examples/stability-ai-out-0.png" "an image of a red square, a blue square and a yellow square" 

We can also execute the image evaluator on batch to process a given folder with images to extract the cosine similarity and aesthetic evaluation of the input:

python image_evaluator.py <mps|cuda|cpu> <image_folder_to_process> <text prompt>

E.g.

python image_evaluator_batch.py mps image_examples "an image of a red square, a blue square and a yellow square"

Reproducing the Tutorial

After the installation, we start by running the EML-gp approach to generate a run to evolve an image aligned with the following prompt: "a red square, a blue square and a yellow square"

python emlart_gp.py 11 1 30 "an image of a red square, a blue square and a yellow square"

The folder "runs" will hold the results of this step. E.g. "runs/emlart-gp/emlart-gp__run__XXXX/00031_ind00000_best_overall.png" will be the best generate image found by running the above command.

We have examples of image generated by some of the latest generative machine learning models on the folder: "image_examples/". We have used the following models for this step:

• Stability AI Stable Diffusion (Replicate)
  https://replicate.com/stability-ai/stable-diffusion?prediction=7cpzt5gmfdrj00cgnv58a6afm0

• Stability AI Stable Diffusion 3.5
  https://huggingface.co/stabilityai/stable-diffusion-3.5-large?text=an+image+of+a+red+square%2C+a+blue+square+and+a+yellow+square

• Stability AI Stable Diffusion XL
  https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0?text=an+image+of+a+red+square+a+blue+square+and+a+yellow+square

• FLUX
  https://huggingface.co/black-forest-labs/FLUX.1-dev?text=an+image+of+a+red+square%2C+a+blue+square+and+a+yellow+square

Generate and download them to the image_examples folder. Next, we can now evaluate our images by using the "image_evaluator_batch.py" script to evaluate and output the results in terms of aesthetic evaluation based on laion aesthetic model and based on cosine similirity between the image features and the prompt.

python image_evaluator_batch.py mps image_examples "an image of a red square, a blue square and a yellow square"

With the image_examples folder of the cloned repository it should yeild the following results:

loading laion model
loading clip model
done!
flux1-dev_2025-1.jpeg - Aesthetic: 5.0050, Similarity: 0.3054
_emlart-gp-00031_ind00000_best_overall-version3.png - Aesthetic: 4.3801, Similarity: 0.3796
_emlart-gp-00031_ind00000_best_overall-version1.png - Aesthetic: 3.9809, Similarity: 0.3738
stability-ai-2024-3.png - Aesthetic: 4.8175, Similarity: 0.2402
stability-ai-2024-2.png - Aesthetic: 4.3327, Similarity: 0.2571
stability-ai-2024-0.png - Aesthetic: 4.9255, Similarity: 0.2888
stable-diffusion-35-large-2025-0.jpeg - Aesthetic: 4.6267, Similarity: 0.3303
stable-diffusion-xl-base-2025-1.jpeg - Aesthetic: 5.9451, Similarity: 0.2091
stability-ai-2024-1.png - Aesthetic: 4.2239, Similarity: 0.2808
stable-diffusion-xl-base-2025-0.jpeg - Aesthetic: 5.1240, Similarity: 0.2482
stable-diffusion-35-large-2025-1.jpeg - Aesthetic: 5.4016, Similarity: 0.3274
chatgpt-2025_00_13_54.png - Aesthetic: 5.1800, Similarity: 0.3013
flux1-dev_2025-2.jpeg - Aesthetic: 3.1616, Similarity: 0.3359
DALL·E 2024-07-14 02.49.31.png - Aesthetic: 5.8161, Similarity: 0.2242
Results exported to evaluation_results.csv