BeautyStandards

Sources for "Erasing 'Ugly' from the Internet: Propagation of the Beauty Myth in Text-Image Models"

📁 Repository Structure

BeautyStandards/Code/
│
├── Prompt_template/
│   ├── beauty_taxonomy.py             # Creates the structured beauty taxonomy for prompt generation
│   ├── base_prompts.txt               # Core set of base prompt templates
│   ├── prompts_metadata.json          # Metadata associated with taxonomy prompts
│   ├── llama3.1_prompts.txt           # Prompts for LLaMA 3.1 model
│   ├── llama3.1_instruct_prompts.txt  # Prompts for LLaMA 3.1 Instruct model
│   ├── deepseek_llm_prompts.txt       # Prompts for DeepSeek LLM
│   └── image_prompts.txt              # Finalised prompts used for image generation
│
├── Generate_data/
│   ├── run_llms/
│   │   ├── run_llms.py                # Runs LLMs with taxonomy-based prompts
│   │   ├── compute_selfBLEU.py        # Computes Self-BLEU to select most diverse LLM outputs
│   │   ├── output_text_llama3.1.txt   # LLaMA 3.1 generated text outputs
│   │   ├── output_text_llama3.1_instruct.txt
│   │   ├── output_text_deepseek_llm.txt # DeepSeek LLM outputs
│   │   └── output_text_selfBleu.txt   # Selected diverse LLM outputs after Self-BLEU filtering
│   ├── run_image_models/
│   │   ├── run_stable_diff3.5_full.py # Generates images via Stable Diffusion 3.5
│   │   ├── run_freepik.py             # Generates images via Freepik API
│   │   └── image_prompts.txt          # Input prompts for image generation
│
├── Results/
│   ├── analyse.ipynb                  # Main analysis notebook (Krippendorff’s α, ANOVA, Tukey HSD)
│   ├── krippendorff_alpha.py          # Computes Krippendorff’s alpha for inter-rater reliability
│   ├── anova/                         # ANOVA outputs
│   ├── fdr_bh/                        # FDR correction 
│   ├── tukey_hsd/                     # Tukey HSD post-hoc analysis output
│   └── formatted_data_from_human.tsv  # Annotated dataset with anonymised Prolific IDs
│
├── LICENSE
└── README.md

⚙️ How to Run the Code

1️⃣ Generate Prompts Using the Beauty Taxonomy

Run the beauty_taxonomy.py script to produce prompts based on diverse beauty-related traits.

cd Code/Prompt_template
python beauty_taxonomy.py

This will create prompt files (e.g., image_prompts.txt, llama3.1_prompts.txt, etc.) that will be used in the next stages. Copy the image_prompts.txt into the run_image_models folder, and other prompt .txt files llama3.1, llama3.1_instruct, deepseek into the run_llms folder.

2️⃣ Generate Text Outputs with LLMs

Move to the run_llms folder and run the language model generation pipeline.

cd ../Generate_data/run_llms
python run_llms.py

This will produce raw text outputs for each model:

• output_text_llama3.1.txt — LLaMA 3.1 outputs
• output_text_llama3.1_instruct.txt — LLaMA 3.1 Instruct outputs
• output_text_deepseek_llm.txt — DeepSeek LLM outputs

🧩 Select the Most Diverse Outputs

Use the Self-BLEU script to identify the most diverse LLM responses for subsequent image generation:

python compute_selfBLEU.py

The filtered outputs will be written to output_text_selfBLEu.txt.

3️⃣ Generate Images

Use both pipelines to produce AI-generated images:

Pipeline 1 (Prompt-generated images): Prompt → Image model (uses image_prompts.txt as prompt input)

Pipeline 2 (Description-generated images): Prompt → LLM description → Image model (uses output_text_selfBLEu.txt as prompt input)

Navigate to run_image_models and run the scripts:

cd ../run_image_models
python run_stable_diff3.5_full.py
python run_freepik.py

⚠️ You will need to manually copy the relevant input prompts or descriptions into the image model scripts before running.

4️⃣ Analyse Results

The Code/Results folder runs all the statistical analysis given in the paper, including anova results and inter rater agreement.

Option 1: Run the Notebook

jupyter notebook analyse.ipynb

The file formatted_data_from_human.tsv contains anonymised human evaluation data.

📜 Citation

⚠️ If you are interested in the image dataset, please contact I.Konstas [at] hw.ac.uk

If you use this code or data, please cite the corresponding paper:

Erasing 'Ugly' from the Internet: Propagation of the Beauty Myth in Text-Image Models
Tanvi Dinkar, Aiqi Jiang, Gavin Abercrombie, Ioannis Konstas
arXiv preprint arXiv:2511.00749, 2025