app.py

import gradio as gr
import torch
from diffusers import (
    DiffusionPipeline,
    StableDiffusionPipeline,
    StableDiffusionXLPipeline,
    EulerDiscreteScheduler,
    UNet2DConditionModel,
    StableDiffusion3Pipeline
)
from transformers import BlipProcessor, BlipForConditionalGeneration
from pathlib import Path
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
from PIL import Image
import matplotlib.pyplot as plt
from matplotlib.colors import hex2color
import stone
import os
import spaces

access_token = os.getenv("AccessTokenSD3")


from huggingface_hub import login
login(token = access_token)


# Define model initialization functions
def load_model(model_name):
    if model_name == "sinteticoXL":
        pipeline = StableDiffusionXLPipeline.from_single_file(
            "https://huggingface.co/lucianosb/sinteticoXL-models/blob/main/sinteticoXL_v1dot2.safetensors",
            torch_dtype=torch.float16,
            variant="fp16",
            use_safetensors=True,
        ).to("cuda")
    elif model_name == "sinteticoXL_Prude":
        pipeline = StableDiffusionXLPipeline.from_single_file(
            "https://huggingface.co/lucianosb/sinteticoXL-models/blob/main/sinteticoXL_prude_v1dot2.safetensors",
            torch_dtype=torch.float16,
            variant="fp16",
            use_safetensors=True,
        ).to("cuda")
    else:
        raise ValueError("Unknown model name")
    return pipeline

# Initialize the default model
default_model = "sinteticoXL"
pipeline_text2image = load_model(default_model)

@spaces.GPU
def getimgen(prompt, model_name):
    if model_name == "sinteticoXL":
        return pipeline_text2image(prompt=prompt, guidance_scale=6.0, num_inference_steps=20).images[0]
    elif model_name == "sinteticoXL_Prude":
        return pipeline_text2image(prompt=prompt, guidance_scale=6.0, num_inference_steps=20).images[0]

blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large")
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large", torch_dtype=torch.float16).to("cuda")

@spaces.GPU
def blip_caption_image(image, prefix):
    inputs = blip_processor(image, prefix, return_tensors="pt").to("cuda", torch.float16)
    out = blip_model.generate(**inputs)
    return blip_processor.decode(out[0], skip_special_tokens=True)

def genderfromcaption(caption):
    cc = caption.split()
    if "man" in cc or "boy" in cc:
        return "Man"
    elif "woman" in cc or "girl" in cc:
        return "Woman"
    return "Unsure"

def genderplot(genlist):    
    order = ["Man", "Woman", "Unsure"]
    words = sorted(genlist, key=lambda x: order.index(x))
    colors = {"Man": "lightgreen", "Woman": "darkgreen", "Unsure": "lightgrey"}
    word_colors = [colors[word] for word in words]
    fig, axes = plt.subplots(2, 5, figsize=(5,5))
    plt.subplots_adjust(hspace=0.1, wspace=0.1)
    for i, ax in enumerate(axes.flat):
        ax.set_axis_off()
        ax.add_patch(plt.Rectangle((0, 0), 1, 1, color=word_colors[i]))
    return fig

def skintoneplot(hex_codes):
    hex_codes = [code for code in hex_codes if code is not None]
    rgb_values = [hex2color(hex_code) for hex_code in hex_codes]
    luminance_values = [0.299 * r + 0.587 * g + 0.114 * b for r, g, b in rgb_values]
    sorted_hex_codes = [code for _, code in sorted(zip(luminance_values, hex_codes), reverse=True)]
    fig, axes = plt.subplots(2, 5, figsize=(5,5))
    plt.subplots_adjust(hspace=0.1, wspace=0.1)
    for i, ax in enumerate(axes.flat):
        ax.set_axis_off()
        if i < len(sorted_hex_codes):
            ax.add_patch(plt.Rectangle((0, 0), 1, 1, color=sorted_hex_codes[i]))
    return fig

@spaces.GPU(duration=200)
def generate_images_plots(prompt, model_name):
    global pipeline_text2image
    pipeline_text2image = load_model(model_name)
    foldername = "temp"
    Path(foldername).mkdir(parents=True, exist_ok=True)
    images = [getimgen(prompt, model_name) for _ in range(10)]
    genders = []
    skintones = []
    for image, i in zip(images, range(10)):
        prompt_prefix = "photo of a "
        caption = blip_caption_image(image, prefix=prompt_prefix)
        image.save(f"{foldername}/image_{i}.png")
        try:
            skintoneres = stone.process(f"{foldername}/image_{i}.png", return_report_image=False)
            tone = skintoneres['faces'][0]['dominant_colors'][0]['color']
            skintones.append(tone)
        except:
            skintones.append(None)
        genders.append(genderfromcaption(caption))
    return images, skintoneplot(skintones), genderplot(genders)

with gr.Blocks(title="Skin Tone and Gender bias in Text-to-Image Generation Models") as demo:
    gr.Markdown("# Skin Tone and Gender bias in Text to Image Models")
    gr.Markdown('''
In this demo, we explore the potential biases in text-to-image models by generating multiple images based on user prompts and analyzing the gender and skin tone of the generated subjects. Here's how the analysis works:

1. **Image Generation**: For each prompt, 10 images are generated using the selected model.
2. **Gender Detection**: The [BLIP caption generator](https://huggingface.co/Salesforce/blip-image-captioning-large) is used to elicit gender markers by identifying words like "man," "boy," "woman," and "girl" in the captions.
3. **Skin Tone Classification**: The [skin-tone-classifier library](https://github.com/ChenglongMa/SkinToneClassifier) is used to extract the skin tones of the generated subjects.


#### Visualization

We create visual grids to represent the data:

- **Skin Tone Grids**: Skin tones are plotted as exact hex codes rather than using the Fitzpatrick scale, which can be [problematic and limiting for darker skin tones](https://arxiv.org/pdf/2309.05148).
- **Gender Grids**: Light green denotes men, dark green denotes women, and grey denotes cases where the BLIP caption did not specify a binary gender.

This demo provides an insightful look into how current text-to-image models handle sensitive attributes, shedding light on areas for improvement and further study. 
[Here is an article](https://medium.com/@evijit/analysis-of-ai-generated-images-of-indian-people-for-colorism-and-sexism-b80ff946759f) showing how this space can be used to perform such analyses, using colorism and sexism in India as an example.
''')
    model_dropdown = gr.Dropdown(
        label="Choose a model", 
        choices=[
            "sinteticoXL",
            "sinteticoXL_Prude"
        ], 
        value=default_model
    )
    prompt = gr.Textbox(label="Enter the Prompt", value = "photo of a beautiful Brazilian woman, high quality, good lighting")
    gallery = gr.Gallery(
        label="Generated images", 
        show_label=False, 
        elem_id="gallery", 
        columns=[5], 
        rows=[2], 
        object_fit="contain", 
        height="auto"
    )
    btn = gr.Button("Generate images", scale=0)
    with gr.Row(equal_height=True):
        skinplot = gr.Plot(label="Skin Tone")
        genplot = gr.Plot(label="Gender")
    btn.click(generate_images_plots, inputs=[prompt, model_dropdown], outputs=[gallery, skinplot, genplot])

demo.launch(debug=True)