File size: 4,284 Bytes
2ec2ebd
 
e2d57c7
2ec2ebd
d415ad5
fe9c201
d415ad5
 
 
 
 
 
7eec65f
d415ad5
 
 
 
 
 
 
e2b5548
 
d415ad5
 
 
 
 
 
3bbb0a9
e2b5548
 
 
 
 
 
 
 
 
 
 
 
d415ad5
 
e2b5548
 
d415ad5
 
 
e2b5548
 
d415ad5
 
 
 
 
dcdc468
d415ad5
e2d57c7
d415ad5
740a0ae
c3eb335
d415ad5
7eec65f
d415ad5
 
dcdc468
740a0ae
2ec2ebd
e18f8f6
d415ad5
 
 
 
 
d329d58
 
d415ad5
 
 
 
 
 
 
 
7eec65f
d415ad5
 
 
 
 
 
 
 
 
 
 
 
edb9b46
d415ad5
edb9b46
d415ad5
 
 
 
 
 
 
 
 
6a0b597
d415ad5
df0a063
f806fdd
824b3c4
f806fdd
824b3c4
6a0b597
 
d415ad5
304f5cd
3bbb0a9
304f5cd
3bbb0a9
 
304f5cd
 
 
 
 
 
 
 
c3eb335
d415ad5
c3eb335
3e50201
 
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
137
138
import gradio as gr
from diffusion_lens import get_images
import numpy as np

MAX_SEED = np.iinfo(np.int32).max

# Description
title = r"""
<h1 align="center">Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines</h1>
"""

description = r"""
<b>A demo for the paper <a href='https://arxiv.org/abs/2403.05846' target='_blank'>Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines</a>.<br>
"""

article = r"""
---
πŸ“ **Citation**
<br>
If our work is helpful for your research or applications, please cite us via:
```
bibtex
@article{toker2024diffusion,
  title={Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines},
  author={Toker, Michael and Orgad, Hadas and Ventura, Mor and Arad, Dana and Belinkov, Yonatan},
  journal={arXiv preprint arXiv:2403.05846},
  year={2024}
}
```
πŸ“§ **Abstact**
<br>
Text-to-image diffusion models (T2I) use a latent representation of a text prompt to guide the image generation process. 
However, the process by which the encoder produces the text representation is unknown. 
We propose the Diffusion Lens, a method for analyzing the text encoder of T2I models by generating images from its intermediate representations. 
Using the Diffusion Lens, we perform an extensive analysis of two recent T2I models. 
Exploring compound prompts, we find that complex scenes describing multiple objects are composed progressively and more slowly compared to simple scenes; 
Exploring knowledge retrieval, we find that representation of uncommon concepts requires further computation compared to common concepts, 
and that knowledge retrieval is gradual across layers. 
Overall, our findings provide valuable insights into the text encoder component in T2I pipelines.
<br>
```
πŸ“§ **Contact**
<br>
If you have any questions, please feel free to open an issue or directly reach us out at <b>tok@cs.technion.ac.il
</b>.
"""




model_num_of_layers = {
    'Stable Diffusion 1.4': 12,
    'Stable Diffusion 2.1': 22,
}


def generate_images(prompt, model, seed):
    seed = random.randint(0, MAX_SEED) if seed == -1 else seed
    print('calling diffusion lens with model:', model, 'and seed:', seed)
    gr.Info('Generating images from intermediate layers..')
    all_images = []  # Initialize a list to store all images
    max_num_of_layers = model_num_of_layers[model]
    for skip_layers in range(max_num_of_layers - 1, -1, -1):
        # Pass the model and seed to the get_images function
        images = get_images(prompt, skip_layers=skip_layers, model=model, seed=seed)
        all_images.append((images[0], f'layer_{max_num_of_layers - skip_layers}'))
        yield all_images

with gr.Blocks() as demo:
    
    gr.Markdown(title)
    gr.Markdown(description)
    
    # text_input = gr.Textbox(label="Enter prompt")
    # model_select = gr.Dropdown(label="Select Model", choices=['sd1', 'sd2'])
    # seed_input = gr.Number(label="Enter Seed", value=0)  # Default seed set to 0
    # Update the submit function to include the new inputs

    
    # text_input.submit(fn=generate_images, inputs=[text_input, model_select, seed_input], outputs=gallery)

    with gr.Column():
        prompt = gr.Textbox(
            label="Prompt",
            value="A photo of Steve Jobs",
        )

    model = gr.Radio(
        [
            "Stable Diffusion 1.4",
            "Stable Diffusion 2.1",
        ],
        value="Stable Diffusion 1.4",
        label="Model",
    )
    
    seed = gr.Slider(
        minimum=0,
        maximum=MAX_SEED,
        value=42,
        step=1,
        label="Seed Value",
    )

    inputs = [
        prompt,
        model,
        seed,
    ]


    generate_button = gr.Button("Generate Image")

    with gr.Column():
        gallery = gr.Gallery(label="Generated Images", columns=6, rows=1, object_fit="contain", height="auto")

    outputs = [gallery]

    gr.on(
        triggers=[
            # prompt.submit,
            generate_button.click,
            # seed.input,
            # model.input
        ],
        fn=generate_images,
        inputs=inputs,
        outputs=outputs,
        show_progress="full",
        show_api=False,
        trigger_mode="always_last",
        )

    gr.Markdown(article)

demo.queue(api_open=False)
demo.launch(show_api=False)