Spaces:
Running
on
A100
Running
on
A100
Update app.py
Browse files
app.py
CHANGED
@@ -6,9 +6,10 @@ from diffusers import (
|
|
6 |
StableDiffusionXLPipeline,
|
7 |
EulerDiscreteScheduler,
|
8 |
UNet2DConditionModel,
|
9 |
-
StableDiffusion3Pipeline
|
|
|
10 |
)
|
11 |
-
from transformers import BlipProcessor, BlipForConditionalGeneration
|
12 |
from pathlib import Path
|
13 |
from safetensors.torch import load_file
|
14 |
from huggingface_hub import hf_hub_download
|
@@ -21,11 +22,9 @@ import spaces
|
|
21 |
|
22 |
access_token = os.getenv("AccessTokenSD3")
|
23 |
|
24 |
-
|
25 |
from huggingface_hub import login
|
26 |
login(token = access_token)
|
27 |
|
28 |
-
|
29 |
# Define model initialization functions
|
30 |
def load_model(model_name):
|
31 |
if model_name == "stabilityai/sdxl-turbo":
|
@@ -34,11 +33,6 @@ def load_model(model_name):
|
|
34 |
torch_dtype=torch.float16,
|
35 |
variant="fp16"
|
36 |
).to("cuda")
|
37 |
-
elif model_name == "runwayml/stable-diffusion-v1-5":
|
38 |
-
pipeline = StableDiffusionPipeline.from_pretrained(
|
39 |
-
model_name,
|
40 |
-
torch_dtype=torch.float16
|
41 |
-
).to("cuda")
|
42 |
elif model_name == "ByteDance/SDXL-Lightning":
|
43 |
base = "stabilityai/stable-diffusion-xl-base-1.0"
|
44 |
ckpt = "sdxl_lightning_4step_unet.safetensors"
|
@@ -70,29 +64,40 @@ def load_model(model_name):
|
|
70 |
scheduler=scheduler,
|
71 |
torch_dtype=torch.float16
|
72 |
).to("cuda")
|
|
|
|
|
|
|
73 |
else:
|
74 |
raise ValueError("Unknown model name")
|
75 |
return pipeline
|
76 |
|
77 |
# Initialize the default model
|
78 |
-
default_model = "
|
79 |
pipeline_text2image = load_model(default_model)
|
80 |
|
81 |
@spaces.GPU
|
82 |
def getimgen(prompt, model_name):
|
83 |
if model_name == "stabilityai/sdxl-turbo":
|
84 |
-
return pipeline_text2image(prompt=prompt, guidance_scale=0.0, num_inference_steps=2).images[0]
|
85 |
-
elif model_name == "runwayml/stable-diffusion-v1-5":
|
86 |
-
return pipeline_text2image(prompt).images[0]
|
87 |
elif model_name == "ByteDance/SDXL-Lightning":
|
88 |
-
return pipeline_text2image(prompt, num_inference_steps=4, guidance_scale=0).images[0]
|
89 |
elif model_name == "segmind/SSD-1B":
|
90 |
neg_prompt = "ugly, blurry, poor quality"
|
91 |
-
return pipeline_text2image(prompt=prompt, negative_prompt=neg_prompt).images[0]
|
92 |
elif model_name == "stabilityai/stable-diffusion-3-medium-diffusers":
|
93 |
-
return pipeline_text2image(prompt=prompt, negative_prompt="", num_inference_steps=28, guidance_scale=7.0).images[0]
|
94 |
elif model_name == "stabilityai/stable-diffusion-2":
|
95 |
-
return pipeline_text2image(prompt=prompt).images[0]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
|
97 |
blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large")
|
98 |
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large", torch_dtype=torch.float16).to("cuda")
|
@@ -136,6 +141,42 @@ def skintoneplot(hex_codes):
|
|
136 |
ax.add_patch(plt.Rectangle((0, 0), 1, 1, color=sorted_hex_codes[i]))
|
137 |
return fig
|
138 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
139 |
@spaces.GPU(duration=200)
|
140 |
def generate_images_plots(prompt, model_name):
|
141 |
global pipeline_text2image
|
@@ -145,6 +186,8 @@ def generate_images_plots(prompt, model_name):
|
|
145 |
images = [getimgen(prompt, model_name) for _ in range(10)]
|
146 |
genders = []
|
147 |
skintones = []
|
|
|
|
|
148 |
for image, i in zip(images, range(10)):
|
149 |
prompt_prefix = "photo of a "
|
150 |
caption = blip_caption_image(image, prefix=prompt_prefix)
|
@@ -156,37 +199,38 @@ def generate_images_plots(prompt, model_name):
|
|
156 |
except:
|
157 |
skintones.append(None)
|
158 |
genders.append(genderfromcaption(caption))
|
159 |
-
|
|
|
|
|
160 |
|
161 |
-
with gr.Blocks(title="
|
162 |
-
gr.Markdown("#
|
163 |
gr.Markdown('''
|
164 |
-
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
|
165 |
-
|
166 |
1. **Image Generation**: For each prompt, 10 images are generated using the selected model.
|
167 |
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.
|
168 |
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.
|
169 |
-
|
170 |
-
|
171 |
#### Visualization
|
172 |
-
|
173 |
We create visual grids to represent the data:
|
174 |
-
|
175 |
- **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).
|
176 |
- **Gender Grids**: Light green denotes men, dark green denotes women, and grey denotes cases where the BLIP caption did not specify a binary gender.
|
177 |
-
|
|
|
|
|
178 |
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.
|
179 |
[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.
|
180 |
''')
|
181 |
model_dropdown = gr.Dropdown(
|
182 |
label="Choose a model",
|
183 |
choices=[
|
|
|
184 |
"stabilityai/stable-diffusion-3-medium-diffusers",
|
185 |
"stabilityai/sdxl-turbo",
|
186 |
"ByteDance/SDXL-Lightning",
|
187 |
"stabilityai/stable-diffusion-2",
|
188 |
-
"
|
189 |
-
"segmind/SSD-1B"
|
190 |
],
|
191 |
value=default_model
|
192 |
)
|
@@ -204,6 +248,9 @@ This demo provides an insightful look into how current text-to-image models hand
|
|
204 |
with gr.Row(equal_height=True):
|
205 |
skinplot = gr.Plot(label="Skin Tone")
|
206 |
genplot = gr.Plot(label="Gender")
|
207 |
-
|
|
|
|
|
|
|
208 |
|
209 |
demo.launch(debug=True)
|
|
|
6 |
StableDiffusionXLPipeline,
|
7 |
EulerDiscreteScheduler,
|
8 |
UNet2DConditionModel,
|
9 |
+
StableDiffusion3Pipeline,
|
10 |
+
FluxPipeline
|
11 |
)
|
12 |
+
from transformers import BlipProcessor, BlipForConditionalGeneration, pipeline
|
13 |
from pathlib import Path
|
14 |
from safetensors.torch import load_file
|
15 |
from huggingface_hub import hf_hub_download
|
|
|
22 |
|
23 |
access_token = os.getenv("AccessTokenSD3")
|
24 |
|
|
|
25 |
from huggingface_hub import login
|
26 |
login(token = access_token)
|
27 |
|
|
|
28 |
# Define model initialization functions
|
29 |
def load_model(model_name):
|
30 |
if model_name == "stabilityai/sdxl-turbo":
|
|
|
33 |
torch_dtype=torch.float16,
|
34 |
variant="fp16"
|
35 |
).to("cuda")
|
|
|
|
|
|
|
|
|
|
|
36 |
elif model_name == "ByteDance/SDXL-Lightning":
|
37 |
base = "stabilityai/stable-diffusion-xl-base-1.0"
|
38 |
ckpt = "sdxl_lightning_4step_unet.safetensors"
|
|
|
64 |
scheduler=scheduler,
|
65 |
torch_dtype=torch.float16
|
66 |
).to("cuda")
|
67 |
+
elif model_name == "black-forest-labs/FLUX.1-dev":
|
68 |
+
pipeline = FluxPipeline.from_pretrained(model_name, torch_dtype=torch.bfloat16)
|
69 |
+
pipeline.enable_model_cpu_offload()
|
70 |
else:
|
71 |
raise ValueError("Unknown model name")
|
72 |
return pipeline
|
73 |
|
74 |
# Initialize the default model
|
75 |
+
default_model = "black-forest-labs/FLUX.1-dev"
|
76 |
pipeline_text2image = load_model(default_model)
|
77 |
|
78 |
@spaces.GPU
|
79 |
def getimgen(prompt, model_name):
|
80 |
if model_name == "stabilityai/sdxl-turbo":
|
81 |
+
return pipeline_text2image(prompt=prompt, guidance_scale=0.0, num_inference_steps=2, height=512, width=512).images[0]
|
|
|
|
|
82 |
elif model_name == "ByteDance/SDXL-Lightning":
|
83 |
+
return pipeline_text2image(prompt, num_inference_steps=4, guidance_scale=0, height=512, width=512).images[0]
|
84 |
elif model_name == "segmind/SSD-1B":
|
85 |
neg_prompt = "ugly, blurry, poor quality"
|
86 |
+
return pipeline_text2image(prompt=prompt, negative_prompt=neg_prompt, height=512, width=512).images[0]
|
87 |
elif model_name == "stabilityai/stable-diffusion-3-medium-diffusers":
|
88 |
+
return pipeline_text2image(prompt=prompt, negative_prompt="", num_inference_steps=28, guidance_scale=7.0, height=512, width=512).images[0]
|
89 |
elif model_name == "stabilityai/stable-diffusion-2":
|
90 |
+
return pipeline_text2image(prompt=prompt, height=512, width=512).images[0]
|
91 |
+
elif model_name == "black-forest-labs/FLUX.1-dev":
|
92 |
+
return pipeline_text2image(
|
93 |
+
prompt,
|
94 |
+
height=512,
|
95 |
+
width=512,
|
96 |
+
guidance_scale=3.5,
|
97 |
+
num_inference_steps=50,
|
98 |
+
max_sequence_length=512,
|
99 |
+
generator=torch.Generator("cpu").manual_seed(0)
|
100 |
+
).images[0]
|
101 |
|
102 |
blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large")
|
103 |
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large", torch_dtype=torch.float16).to("cuda")
|
|
|
141 |
ax.add_patch(plt.Rectangle((0, 0), 1, 1, color=sorted_hex_codes[i]))
|
142 |
return fig
|
143 |
|
144 |
+
def age_detector(image):
|
145 |
+
pipe = pipeline('image-classification', model="dima806/faces_age_detection", device=0)
|
146 |
+
result = pipe(image)
|
147 |
+
max_score_item = max(result, key=lambda item: item['score'])
|
148 |
+
return max_score_item['label']
|
149 |
+
|
150 |
+
def ageplot(agelist):
|
151 |
+
order = ["YOUNG", "MIDDLE", "OLD"]
|
152 |
+
words = sorted(agelist, key=lambda x: order.index(x))
|
153 |
+
colors = {"YOUNG": "skyblue", "MIDDLE": "royalblue", "OLD": "darkblue"}
|
154 |
+
word_colors = [colors[word] for word in words]
|
155 |
+
fig, axes = plt.subplots(2, 5, figsize=(5,5))
|
156 |
+
plt.subplots_adjust(hspace=0.1, wspace=0.1)
|
157 |
+
for i, ax in enumerate(axes.flat):
|
158 |
+
ax.set_axis_off()
|
159 |
+
ax.add_patch(plt.Rectangle((0, 0), 1, 1, color=word_colors[i]))
|
160 |
+
return fig
|
161 |
+
|
162 |
+
def is_nsfw(image):
|
163 |
+
classifier = pipeline("image-classification", model="Falconsai/nsfw_image_detection")
|
164 |
+
result = classifier(image)
|
165 |
+
max_score_item = max(result, key=lambda item: item['score'])
|
166 |
+
return max_score_item['label']
|
167 |
+
|
168 |
+
def nsfwplot(nsfwlist):
|
169 |
+
order = ["normal", "nsfw"]
|
170 |
+
words = sorted(nsfwlist, key=lambda x: order.index(x))
|
171 |
+
colors = {"normal": "mistyrose", "nsfw": "red"}
|
172 |
+
word_colors = [colors[word] for word in words]
|
173 |
+
fig, axes = plt.subplots(2, 5, figsize=(5,5))
|
174 |
+
plt.subplots_adjust(hspace=0.1, wspace=0.1)
|
175 |
+
for i, ax in enumerate(axes.flat):
|
176 |
+
ax.set_axis_off()
|
177 |
+
ax.add_patch(plt.Rectangle((0, 0), 1, 1, color=word_colors[i]))
|
178 |
+
return fig
|
179 |
+
|
180 |
@spaces.GPU(duration=200)
|
181 |
def generate_images_plots(prompt, model_name):
|
182 |
global pipeline_text2image
|
|
|
186 |
images = [getimgen(prompt, model_name) for _ in range(10)]
|
187 |
genders = []
|
188 |
skintones = []
|
189 |
+
ages = []
|
190 |
+
nsfws = []
|
191 |
for image, i in zip(images, range(10)):
|
192 |
prompt_prefix = "photo of a "
|
193 |
caption = blip_caption_image(image, prefix=prompt_prefix)
|
|
|
199 |
except:
|
200 |
skintones.append(None)
|
201 |
genders.append(genderfromcaption(caption))
|
202 |
+
ages.append(age_detector(image))
|
203 |
+
nsfws.append(is_nsfw(image))
|
204 |
+
return images, skintoneplot(skintones), genderplot(genders), ageplot(ages), nsfwplot(nsfws)
|
205 |
|
206 |
+
with gr.Blocks(title="Demographic bias in Text-to-Image Generation Models") as demo:
|
207 |
+
gr.Markdown("# Demographic bias in Text to Image Models")
|
208 |
gr.Markdown('''
|
209 |
+
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:
|
|
|
210 |
1. **Image Generation**: For each prompt, 10 images are generated using the selected model.
|
211 |
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.
|
212 |
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.
|
213 |
+
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.
|
214 |
+
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).
|
215 |
#### Visualization
|
|
|
216 |
We create visual grids to represent the data:
|
|
|
217 |
- **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).
|
218 |
- **Gender Grids**: Light green denotes men, dark green denotes women, and grey denotes cases where the BLIP caption did not specify a binary gender.
|
219 |
+
- **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.
|
220 |
+
- **NFAA Grids**: Light red denotes FAA images, and dark red denotes NFAA images.
|
221 |
+
|
222 |
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.
|
223 |
[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.
|
224 |
''')
|
225 |
model_dropdown = gr.Dropdown(
|
226 |
label="Choose a model",
|
227 |
choices=[
|
228 |
+
"black-forest-labs/FLUX.1-dev",
|
229 |
"stabilityai/stable-diffusion-3-medium-diffusers",
|
230 |
"stabilityai/sdxl-turbo",
|
231 |
"ByteDance/SDXL-Lightning",
|
232 |
"stabilityai/stable-diffusion-2",
|
233 |
+
"segmind/SSD-1B",
|
|
|
234 |
],
|
235 |
value=default_model
|
236 |
)
|
|
|
248 |
with gr.Row(equal_height=True):
|
249 |
skinplot = gr.Plot(label="Skin Tone")
|
250 |
genplot = gr.Plot(label="Gender")
|
251 |
+
with gr.Row(equal_height=True):
|
252 |
+
agesplot = gr.Plot(label="Age")
|
253 |
+
nsfwsplot = gr.Plot(label="NFAA")
|
254 |
+
btn.click(generate_images_plots, inputs=[prompt, model_dropdown], outputs=[gallery, skinplot, genplot, agesplot, nsfwsplot])
|
255 |
|
256 |
demo.launch(debug=True)
|