File size: 5,045 Bytes
e8994f0
 
 
 
 
 
 
 
 
12ec423
e8994f0
 
e8de545
e8994f0
 
 
 
 
 
 
02e0d06
 
e8994f0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
02e0d06
 
fe957af
12ec423
829a688
 
 
 
 
 
 
 
 
e8994f0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fe957af
 
e8994f0
e8de545
 
 
 
 
02e0d06
12ec423
 
 
 
e8994f0
e8de545
829a688
e8994f0
 
 
12ec423
 
 
 
 
 
 
e8994f0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
829a688
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
import gradio as gr
from PIL import Image as PILImage
import tempfile

from modules.ddg_search import search_images
from modules.embeddings import get_embedding, load_image_from_url
from modules.similarity import compute_similarity
from modules.pdf_report import create_pdf_report


def analyze_image_stream(prompt, uploaded_img, num_results):
    if not uploaded_img or not prompt:
        yield None, None, None, "❌ Please provide both a prompt and an image.", *[gr.update(visible=False)] * 6
        return

    user_img = uploaded_img.convert("RGB")
    urls = search_images(prompt, max_results=num_results)
    retrieved_images = [(url, "") for url in urls]

    yield retrieved_images, None, None, "πŸ”„ Computing similarity...", \
          gr.update(visible=True), gr.update(visible=False), gr.update(visible=False), \
          gr.update(visible=True), gr.update(visible=False), gr.update(visible=False)

    user_emb = get_embedding(user_img)
    results = []
    for url in urls:
        try:
            img = load_image_from_url(url)
            emb = get_embedding(img)
            sim = compute_similarity(user_emb, emb)
            results.append((url, sim))
        except:
            continue

    results = sorted(results, key=lambda x: x[1], reverse=True)
    top_results = results[:5]
    top_images = [(url, f"Similarity: {sim:.3f}") for url, sim in top_results]

    pdf_bytes = create_pdf_report(prompt, user_img, top_results)
    with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf") as f:
        f.write(pdf_bytes)
        pdf_path = f.name

    yield retrieved_images, top_images, pdf_path, "βœ… Done!", \
          gr.update(visible=True), gr.update(visible=True), gr.update(visible=True), \
          gr.update(visible=True), gr.update(visible=True), gr.update(visible=True)


with gr.Blocks(title="Influence Analyzer Demo", css="""
#uploaded_image img {
    object-fit: contain;
    height: 250px;
    width: auto;
    border: 1px solid #ccc;
    margin-top: 5px;
}
""") as demo:
    gr.Markdown("""
### πŸ” Search-Based Influence Analyzer Demo
This tool supports the **interpretability of generative black-box models** by identifying potentially influential training data.

**How it works:**
- πŸ”Ž Retrieves public images using DuckDuckGo based on your prompt.
- 🧠 Computes embeddings and compares them to your uploaded image to find the most similar results.

The goal is to provide a proxy for understanding which real-world images may have contributed to a generation β€” useful for **interpretability**, **data attribution**, and **copyright assessment**.
""")

    gr.Markdown("### πŸ“ Provide Your Prompt and Image")

    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(label="Prompt", placeholder="Describe your image prompt...", lines=1)
            num_results = gr.Slider(5, 50, value=30, step=1, label="Images to retrieve")

            gr.Markdown("**✏️ Try an example**")
            examples = [
                ("a horse in the moon", "examples/moon_horse.png"),
                ("a greek philosopher drinking beer", "examples/greek_beer.png"),
            ]

            example_btns = []
            for i, (text, path) in enumerate(examples):
                btn = gr.Button(text)
                example_btns.append((btn, text, path))

        with gr.Column():
            uploaded_img = gr.Image(type="pil", label="Upload your generated image", elem_id="uploaded_image")

    run_btn = gr.Button("πŸš€ Run Search & Compare", variant="primary")

    for btn, prompt_text, image_path in example_btns:
        def make_loader(p=prompt_text, img=image_path):
            def _load():
                return p, PILImage.open(img)
            return _load
        btn.click(fn=make_loader(), inputs=[], outputs=[prompt, uploaded_img])

    gr.Markdown("---")
    retrieved_heading = gr.Markdown("### πŸ” Retrieved Images", visible=False)
    retrieved_gallery = gr.Gallery(label="", columns=5, height="auto", visible=False)

    similar_heading = gr.Markdown("### βœ… Top 5 Most Similar Images", visible=False)
    similar_gallery = gr.Gallery(label="", columns=5, height="auto", visible=False)

    output_heading = gr.Markdown("### πŸ“„ Output Report", visible=False)
    pdf_out = gr.File(label="Download Similarity Report", visible=False)

    status = gr.Textbox(label="Status", interactive=False)

    run_btn.click(
        fn=analyze_image_stream,
        inputs=[prompt, uploaded_img, num_results],
        outputs=[
            retrieved_gallery,
            similar_gallery,
            pdf_out,
            status,
            retrieved_heading,
            similar_heading,
            output_heading,
            retrieved_gallery,
            similar_gallery,
            pdf_out,
        ],
    )

    gr.Markdown("---")
    gr.Markdown(
        "🧠 *This demo is part of ongoing research in interpretability for generative models. "
        "If you're interested in collaboration, feel free to get in touch.*"
    )

demo.launch()