Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,8 +1,9 @@
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
from transformers import CLIPProcessor, CLIPModel, BlipProcessor, BlipForConditionalGeneration
|
4 |
-
from PIL import Image
|
5 |
import numpy as np
|
|
|
6 |
from openai import OpenAI
|
7 |
|
8 |
# 初始化模型
|
@@ -11,13 +12,77 @@ clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
|
11 |
blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
12 |
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
|
13 |
|
14 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
def analyze_images(image_a, image_b, api_key):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
# BLIP生成描述
|
17 |
-
|
18 |
-
|
19 |
-
caption = blip_model.generate(**inputs)
|
20 |
-
return blip_processor.decode(caption[0], skip_special_tokens=True)
|
21 |
|
22 |
# CLIP特征提取
|
23 |
def extract_features(image):
|
@@ -25,24 +90,11 @@ def analyze_images(image_a, image_b, api_key):
|
|
25 |
features = clip_model.get_image_features(**inputs)
|
26 |
return features.detach().numpy()
|
27 |
|
28 |
-
# 图像已经是 PIL.Image 对象,直接处理
|
29 |
-
img_a = image_a.convert("RGB")
|
30 |
-
img_b = image_b.convert("RGB")
|
31 |
-
|
32 |
-
# 生成描述
|
33 |
-
caption_a = generate_caption(img_a)
|
34 |
-
caption_b = generate_caption(img_b)
|
35 |
-
|
36 |
-
# 提取特征
|
37 |
features_a = extract_features(img_a)
|
38 |
features_b = extract_features(img_b)
|
39 |
-
|
40 |
-
# 计算嵌入相似性
|
41 |
-
cosine_similarity = np.dot(features_a, features_b.T) / (np.linalg.norm(features_a) * np.linalg.norm(features_b))
|
42 |
latent_diff = np.abs(features_a - features_b).tolist()
|
43 |
|
44 |
-
# 调用
|
45 |
-
client = OpenAI(api_key=api_key, base_url="https://api.deepseek.com")
|
46 |
gpt_response = client.chat.completions.create(
|
47 |
model="deepseek-chat",
|
48 |
messages=[
|
@@ -51,45 +103,79 @@ def analyze_images(image_a, image_b, api_key):
|
|
51 |
],
|
52 |
stream=False
|
53 |
)
|
54 |
-
|
55 |
-
|
56 |
-
#
|
|
|
|
|
57 |
return {
|
58 |
"caption_a": caption_a,
|
59 |
"caption_b": caption_b,
|
60 |
-
"
|
61 |
-
"
|
62 |
-
"
|
|
|
63 |
}
|
64 |
|
65 |
-
#
|
66 |
with gr.Blocks() as demo:
|
67 |
-
gr.Markdown("#
|
68 |
-
|
|
|
69 |
with gr.Row():
|
70 |
with gr.Column():
|
71 |
-
image_a = gr.Image(label="图片A", type="pil")
|
72 |
with gr.Column():
|
73 |
-
image_b = gr.Image(label="图片B", type="pil")
|
74 |
-
|
75 |
-
api_key_input = gr.Textbox(label="API Key", placeholder="输入您的 DeepSeek API Key", type="password")
|
76 |
-
|
77 |
analyze_button = gr.Button("分析图片")
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
# 分析逻辑
|
85 |
def process_analysis(img_a, img_b, api_key):
|
86 |
results = analyze_images(img_a, img_b, api_key)
|
87 |
-
|
88 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
analyze_button.click(
|
90 |
fn=process_analysis,
|
91 |
inputs=[image_a, image_b, api_key_input],
|
92 |
-
outputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
)
|
94 |
|
95 |
demo.launch()
|
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
from transformers import CLIPProcessor, CLIPModel, BlipProcessor, BlipForConditionalGeneration
|
4 |
+
from PIL import Image, ImageChops
|
5 |
import numpy as np
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
from openai import OpenAI
|
8 |
|
9 |
# 初始化模型
|
|
|
12 |
blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
13 |
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
|
14 |
|
15 |
+
# 图像处理函数
|
16 |
+
def compute_difference_images(img_a, img_b):
|
17 |
+
# 线稿提取
|
18 |
+
def extract_sketch(image):
|
19 |
+
grayscale = image.convert("L")
|
20 |
+
inverted = ImageChops.invert(grayscale)
|
21 |
+
sketch = ImageChops.screen(grayscale, inverted)
|
22 |
+
return sketch
|
23 |
+
|
24 |
+
# 法向量图像(模拟法向量处理为简单的边缘增强)
|
25 |
+
def compute_normal_map(image):
|
26 |
+
edges = image.filter(ImageFilter.FIND_EDGES)
|
27 |
+
return edges
|
28 |
+
|
29 |
+
# 图像混合差异
|
30 |
+
diff_overlay = ImageChops.difference(img_a, img_b)
|
31 |
+
|
32 |
+
return {
|
33 |
+
"original_a": img_a,
|
34 |
+
"original_b": img_b,
|
35 |
+
"sketch_a": extract_sketch(img_a),
|
36 |
+
"sketch_b": extract_sketch(img_b),
|
37 |
+
"normal_a": compute_normal_map(img_a),
|
38 |
+
"normal_b": compute_normal_map(img_b),
|
39 |
+
"diff_overlay": diff_overlay
|
40 |
+
}
|
41 |
+
|
42 |
+
# BLIP生成更详尽描述
|
43 |
+
def generate_detailed_caption(image):
|
44 |
+
inputs = blip_processor(image, return_tensors="pt")
|
45 |
+
caption = blip_model.generate(**inputs, max_length=128, num_beams=5, no_repeat_ngram_size=2)
|
46 |
+
return blip_processor.decode(caption[0], skip_special_tokens=True)
|
47 |
+
|
48 |
+
# 特征差异可视化
|
49 |
+
def plot_feature_differences(latent_diff):
|
50 |
+
diff_magnitude = [abs(x) for x in latent_diff[0]]
|
51 |
+
indices = range(len(diff_magnitude))
|
52 |
+
|
53 |
+
# 柱状图
|
54 |
+
plt.figure(figsize=(8, 4))
|
55 |
+
plt.bar(indices, diff_magnitude, alpha=0.7)
|
56 |
+
plt.xlabel("Feature Index")
|
57 |
+
plt.ylabel("Magnitude of Difference")
|
58 |
+
plt.title("Feature Differences (Bar Chart)")
|
59 |
+
bar_chart_path = "bar_chart.png"
|
60 |
+
plt.savefig(bar_chart_path)
|
61 |
+
plt.close()
|
62 |
+
|
63 |
+
# 饼图
|
64 |
+
plt.figure(figsize=(6, 6))
|
65 |
+
plt.pie(diff_magnitude[:10], labels=range(10), autopct="%1.1f%%", startangle=140)
|
66 |
+
plt.title("Top 10 Feature Differences (Pie Chart)")
|
67 |
+
pie_chart_path = "pie_chart.png"
|
68 |
+
plt.savefig(pie_chart_path)
|
69 |
+
plt.close()
|
70 |
+
|
71 |
+
return bar_chart_path, pie_chart_path
|
72 |
+
|
73 |
+
# 分析函数
|
74 |
def analyze_images(image_a, image_b, api_key):
|
75 |
+
# 调用 OpenAI 客户端
|
76 |
+
client = OpenAI(api_key=api_key, base_url="https://api.deepseek.com")
|
77 |
+
|
78 |
+
# 图像差异处理
|
79 |
+
img_a = image_a.convert("RGB")
|
80 |
+
img_b = image_b.convert("RGB")
|
81 |
+
images_diff = compute_difference_images(img_a, img_b)
|
82 |
+
|
83 |
# BLIP生成描述
|
84 |
+
caption_a = generate_detailed_caption(img_a)
|
85 |
+
caption_b = generate_detailed_caption(img_b)
|
|
|
|
|
86 |
|
87 |
# CLIP特征提取
|
88 |
def extract_features(image):
|
|
|
90 |
features = clip_model.get_image_features(**inputs)
|
91 |
return features.detach().numpy()
|
92 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
features_a = extract_features(img_a)
|
94 |
features_b = extract_features(img_b)
|
|
|
|
|
|
|
95 |
latent_diff = np.abs(features_a - features_b).tolist()
|
96 |
|
97 |
+
# 调用 GPT 获取更详细描述
|
|
|
98 |
gpt_response = client.chat.completions.create(
|
99 |
model="deepseek-chat",
|
100 |
messages=[
|
|
|
103 |
],
|
104 |
stream=False
|
105 |
)
|
106 |
+
text_analysis = gpt_response.choices[0].message.content.strip()
|
107 |
+
|
108 |
+
# 可视化特征差异
|
109 |
+
bar_chart_path, pie_chart_path = plot_feature_differences(latent_diff)
|
110 |
+
|
111 |
return {
|
112 |
"caption_a": caption_a,
|
113 |
"caption_b": caption_b,
|
114 |
+
"text_analysis": text_analysis,
|
115 |
+
"images_diff": images_diff,
|
116 |
+
"bar_chart": bar_chart_path,
|
117 |
+
"pie_chart": pie_chart_path
|
118 |
}
|
119 |
|
120 |
+
# Gradio界面
|
121 |
with gr.Blocks() as demo:
|
122 |
+
gr.Markdown("# 图像对比分析工具")
|
123 |
+
api_key_input = gr.Textbox(label="API Key", placeholder="输入您的 DeepSeek API Key", type="password")
|
124 |
+
|
125 |
with gr.Row():
|
126 |
with gr.Column():
|
127 |
+
image_a = gr.Image(label="图片A", type="pil")
|
128 |
with gr.Column():
|
129 |
+
image_b = gr.Image(label="图片B", type="pil")
|
130 |
+
|
|
|
|
|
131 |
analyze_button = gr.Button("分析图片")
|
132 |
+
|
133 |
+
with gr.Row():
|
134 |
+
gr.Markdown("## 图像差异")
|
135 |
+
result_diff = gr.Gallery(label="混合差异图像").style(grid=3)
|
136 |
+
|
137 |
+
with gr.Row():
|
138 |
+
result_caption_a = gr.Textbox(label="图片A描述", interactive=False)
|
139 |
+
result_caption_b = gr.Textbox(label="图片B描述", interactive=False)
|
140 |
+
|
141 |
+
with gr.Row():
|
142 |
+
gr.Markdown("## 差异分析")
|
143 |
+
result_text_analysis = gr.Textbox(label="详细分析", interactive=False, lines=5)
|
144 |
+
result_bar_chart = gr.Image(label="特征差异柱状图")
|
145 |
+
result_pie_chart = gr.Image(label="特征差异饼图")
|
146 |
+
|
147 |
# 分析逻辑
|
148 |
def process_analysis(img_a, img_b, api_key):
|
149 |
results = analyze_images(img_a, img_b, api_key)
|
150 |
+
diff_images = [
|
151 |
+
("Original A", results["images_diff"]["original_a"]),
|
152 |
+
("Original B", results["images_diff"]["original_b"]),
|
153 |
+
("Sketch A", results["images_diff"]["sketch_a"]),
|
154 |
+
("Sketch B", results["images_diff"]["sketch_b"]),
|
155 |
+
("Normal A", results["images_diff"]["normal_a"]),
|
156 |
+
("Normal B", results["images_diff"]["normal_b"]),
|
157 |
+
("Difference Overlay", results["images_diff"]["diff_overlay"]),
|
158 |
+
]
|
159 |
+
return (
|
160 |
+
diff_images,
|
161 |
+
results["caption_a"],
|
162 |
+
results["caption_b"],
|
163 |
+
results["text_analysis"],
|
164 |
+
results["bar_chart"],
|
165 |
+
results["pie_chart"]
|
166 |
+
)
|
167 |
+
|
168 |
analyze_button.click(
|
169 |
fn=process_analysis,
|
170 |
inputs=[image_a, image_b, api_key_input],
|
171 |
+
outputs=[
|
172 |
+
result_diff,
|
173 |
+
result_caption_a,
|
174 |
+
result_caption_b,
|
175 |
+
result_text_analysis,
|
176 |
+
result_bar_chart,
|
177 |
+
result_pie_chart
|
178 |
+
]
|
179 |
)
|
180 |
|
181 |
demo.launch()
|