Update app.py

app.py

@@ -6,9 +6,10 @@ from diffusers import (
     StableDiffusionXLPipeline,
     EulerDiscreteScheduler,
     UNet2DConditionModel,
-    StableDiffusion3Pipeline
+    StableDiffusion3Pipeline,
+    FluxPipeline
 )
-from transformers import BlipProcessor, BlipForConditionalGeneration
+from transformers import BlipProcessor, BlipForConditionalGeneration, pipeline
 from pathlib import Path
 from safetensors.torch import load_file
 from huggingface_hub import hf_hub_download
@@ -21,11 +22,9 @@ 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 == "stabilityai/sdxl-turbo":
@@ -34,11 +33,6 @@ def load_model(model_name):
             torch_dtype=torch.float16, 
             variant="fp16"
         ).to("cuda")
-    elif model_name == "runwayml/stable-diffusion-v1-5":
-        pipeline = StableDiffusionPipeline.from_pretrained(
-            model_name, 
-            torch_dtype=torch.float16
-        ).to("cuda")
     elif model_name == "ByteDance/SDXL-Lightning":
         base = "stabilityai/stable-diffusion-xl-base-1.0"
         ckpt = "sdxl_lightning_4step_unet.safetensors"
@@ -70,29 +64,40 @@ def load_model(model_name):
             scheduler=scheduler, 
             torch_dtype=torch.float16
         ).to("cuda")
+    elif model_name == "black-forest-labs/FLUX.1-dev":
+        pipeline = FluxPipeline.from_pretrained(model_name, torch_dtype=torch.bfloat16)
+        pipeline.enable_model_cpu_offload()
     else:
         raise ValueError("Unknown model name")
     return pipeline
 
 # Initialize the default model
-default_model = "stabilityai/stable-diffusion-3-medium-diffusers"
+default_model = "black-forest-labs/FLUX.1-dev"
 pipeline_text2image = load_model(default_model)
 
 @spaces.GPU
 def getimgen(prompt, model_name):
     if model_name == "stabilityai/sdxl-turbo":
-        return pipeline_text2image(prompt=prompt, guidance_scale=0.0, num_inference_steps=2).images[0]
-    elif model_name == "runwayml/stable-diffusion-v1-5":
-        return pipeline_text2image(prompt).images[0]
+        return pipeline_text2image(prompt=prompt, guidance_scale=0.0, num_inference_steps=2, height=512, width=512).images[0]
     elif model_name == "ByteDance/SDXL-Lightning":
-        return pipeline_text2image(prompt, num_inference_steps=4, guidance_scale=0).images[0]
+        return pipeline_text2image(prompt, num_inference_steps=4, guidance_scale=0, height=512, width=512).images[0]
     elif model_name == "segmind/SSD-1B":
         neg_prompt = "ugly, blurry, poor quality"
-        return pipeline_text2image(prompt=prompt, negative_prompt=neg_prompt).images[0]
+        return pipeline_text2image(prompt=prompt, negative_prompt=neg_prompt, height=512, width=512).images[0]
     elif model_name == "stabilityai/stable-diffusion-3-medium-diffusers":
-        return pipeline_text2image(prompt=prompt, negative_prompt="", num_inference_steps=28, guidance_scale=7.0).images[0]
+        return pipeline_text2image(prompt=prompt, negative_prompt="", num_inference_steps=28, guidance_scale=7.0, height=512, width=512).images[0]
     elif model_name == "stabilityai/stable-diffusion-2":
-        return pipeline_text2image(prompt=prompt).images[0]
+        return pipeline_text2image(prompt=prompt, height=512, width=512).images[0]
+    elif model_name == "black-forest-labs/FLUX.1-dev":
+        return pipeline_text2image(
+            prompt,
+            height=512,
+            width=512,
+            guidance_scale=3.5,
+            num_inference_steps=50,
+            max_sequence_length=512,
+            generator=torch.Generator("cpu").manual_seed(0)
+        ).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")
@@ -136,6 +141,42 @@ def skintoneplot(hex_codes):
             ax.add_patch(plt.Rectangle((0, 0), 1, 1, color=sorted_hex_codes[i]))
     return fig
 
+def age_detector(image):
+    pipe = pipeline('image-classification', model="dima806/faces_age_detection", device=0)
+    result = pipe(image)
+    max_score_item = max(result, key=lambda item: item['score'])
+    return max_score_item['label']
+
+def ageplot(agelist):
+    order = ["YOUNG", "MIDDLE", "OLD"]
+    words = sorted(agelist, key=lambda x: order.index(x))
+    colors = {"YOUNG": "skyblue", "MIDDLE": "royalblue", "OLD": "darkblue"}
+    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 is_nsfw(image):
+    classifier = pipeline("image-classification", model="Falconsai/nsfw_image_detection")
+    result = classifier(image)
+    max_score_item = max(result, key=lambda item: item['score'])
+    return max_score_item['label']
+
+def nsfwplot(nsfwlist):
+    order = ["normal", "nsfw"]
+    words = sorted(nsfwlist, key=lambda x: order.index(x))
+    colors = {"normal": "mistyrose", "nsfw": "red"}
+    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
+
 @spaces.GPU(duration=200)
 def generate_images_plots(prompt, model_name):
     global pipeline_text2image
@@ -145,6 +186,8 @@ def generate_images_plots(prompt, model_name):
     images = [getimgen(prompt, model_name) for _ in range(10)]
     genders = []
     skintones = []
+    ages = []
+    nsfws = []
     for image, i in zip(images, range(10)):
         prompt_prefix = "photo of a "
         caption = blip_caption_image(image, prefix=prompt_prefix)
@@ -156,37 +199,38 @@ def generate_images_plots(prompt, model_name):
         except:
             skintones.append(None)
         genders.append(genderfromcaption(caption))
-    return images, skintoneplot(skintones), genderplot(genders)
+        ages.append(age_detector(image))
+        nsfws.append(is_nsfw(image))
+    return images, skintoneplot(skintones), genderplot(genders), ageplot(ages), nsfwplot(nsfws)
 
-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")
+with gr.Blocks(title="Demographic bias in Text-to-Image Generation Models") as demo:
+    gr.Markdown("# Demographic 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:
-
+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, skin tone, age, and potential sexual nature 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.
-
-
+4. **Age Detection**: The [Faces Age Detection model](https://huggingface.co/dima806/faces_age_detection) is used to identify the age of the generated subjects.
+5. **NFAA Detection**: The [Falconsai/nsfw_image_detection](https://huggingface.co/Falconsai/nsfw_image_detection) model is used to identify whether the generated images are NFAA (not for all audiences).
 #### 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.
-
+- **Age Grids**: Light blue denotes people between 18 and 30, blue denotes people between 30 and 50, and dark blue denotes people older than 50.
+- **NFAA Grids**: Light red denotes FAA images, and dark red denotes NFAA images.
+                
 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=[
+            "black-forest-labs/FLUX.1-dev",
             "stabilityai/stable-diffusion-3-medium-diffusers",
             "stabilityai/sdxl-turbo", 
             "ByteDance/SDXL-Lightning",
             "stabilityai/stable-diffusion-2",
-            "runwayml/stable-diffusion-v1-5", 
-            "segmind/SSD-1B"
+            "segmind/SSD-1B",
         ], 
         value=default_model
     )
@@ -204,6 +248,9 @@ This demo provides an insightful look into how current text-to-image models hand
     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])
+    with gr.Row(equal_height=True):
+        agesplot = gr.Plot(label="Age")
+        nsfwsplot = gr.Plot(label="NFAA")
+    btn.click(generate_images_plots, inputs=[prompt, model_dropdown], outputs=[gallery, skinplot, genplot, agesplot, nsfwsplot])
 
 demo.launch(debug=True)