Debiasing Multi-Language Large Vision-Language Models

Context

Large Vision-Language Models (LVLMs) are increasingly popular and influential in e-commerce, retail, content creation, and fashion. Despite their growing utility, these models are known to inherit social biases from their underlying vision and language components. The research aims to address these biases, particularly in the context of non-English-speaking users, and to make LVLMs more accessible and equitable.

Setup

Create a virtual environment

python3 -m venv venv
source ./venv/bin/activate
(venv) pip install -r requirements.txt

Train Data Generation

The first step in fine-tuning the Alt-Diffusion Model is creating the datasets required for training. This involves generating prompts in various languages, creating images based on those prompts, and finally, combining the results into cohesive datasets.

Generate English Prompts

We begin by creating the English prompts using the following pattern:

<STYLE> photo of a <AGE> <GENDER> <PROFESSION> <DRESSING> <PLACE>

To generate these prompts, run the following command:

python -m src.generate_prompts

Generate Translated Prompts

Translate the English prompts into German, Chinese, and Japanese:

python -m src.translate.translate_prompts

Generate Images

Generate images using all prompts:

python -m src.generate_images

Create the datasets

Combine the generated prompts and images to create a dataset for each experiment. You can choose the source of the prompts and the source of the images:

python -m src.collect_datasets

Fine-tune Alt-Diffusion model

To fine-tune the Alt-Diffusion model, run the train_text_to_image CLI.

CLI Parameters:

• —pretrained_model_name_or_path BAAI/AltDiffusion-m9
• —train_data_dir dataset_path
• —mixed_precision fp16
• —allow_tf32
• —gradient_checkpointing
• —resolution 256
• —center_crop
• —random_flip
• —train_batch_size 16
• —gradient_accumulation_steps 2
• —num_train_epochs 5
• —learning_rate 5e-06
• —lr_scheduler polynomial
• —lr_warmup_steps 0
• —output_dir "output_fine_tune/alt_diffusion_m9_fine_tune"

python -m src.tti-fine-tune.train_text_to_image

Evaluation Results

Bias

First, we need to run classify.py CLI to find faces and attributes (e.g. gender, race, etc.)

CLI Parameters:

• —model (default AD)
• —model_version (default m9)
• —classifier (default fairface)
• —dataset (we only use the default value, occupations)
• —language
• —num_images (default 25)
• —postfix. This is a parameter that we added to manipulate paths if we move stuff around. For example, with the base postfix, we evaluate data/eval/generated_images_base/...

python -m src.evaluate.classify --language english

This will extract faces from images generated with generate_images.py and output a file like occupations_fairface_generated.txt with lines like

musician
man: X, woman: Y

To compute the bias metrics from the extracted faces, run the evaluate_bias CLI.

CLI Parameters:

• —model (default AD)
• —model_version (default m9)
• —classifier (default fairface)
• —dataset (we only use the default value, occupations)
• —language

python -m src.evaluate.evaluate_bias --model_version finetune_english --language english

CLIP/Alignment

To evaluate for image/text alignment (CLIP score), run the evaluate_alignment CLI.

CLI Parameters:

• —model (default AD)
• — model_version (default m9)
• —mode (we only use the default value, generated)
• —classifier (default fairface)
• —dataset (we only use the default value, occupations)
• —language
• —num_images (default 25)
• —postfix

python -m src.evaluate.evaluate_alignment --language english --postfix base

Evaluates the images in data/eval/generated_images_base/multilang/AD_m9/english/generated/[..] for all professions and outputs the results in resutls/multilang/AD_m9/english/occupations_generated_clipscore.txt