File size: 11,144 Bytes
4c022fe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3c42b4c
4c022fe
 
 
 
0a499ee
 
 
ba87c8b
0a499ee
4c022fe
 
 
 
 
 
 
 
51be712
4c022fe
 
 
 
 
 
ed9d237
4c022fe
 
 
 
 
 
 
 
 
d0745b6
4c022fe
 
 
 
 
 
 
 
 
 
 
 
ed9d237
4c022fe
 
 
 
 
 
 
 
 
 
 
 
 
ab7db7f
4c022fe
ab7db7f
4c022fe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ab7db7f
 
4c022fe
 
 
0a499ee
 
4c022fe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69e92de
 
 
 
 
4c022fe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ba87c8b
 
 
4c022fe
 
 
bc4d8bb
4c022fe
 
 
 
ed9d237
4c022fe
 
bc4d8bb
4c022fe
 
 
ae61c93
ed9d237
4c022fe
 
ae61c93
4c022fe
 
 
 
ed9d237
4c022fe
fb8cc1d
ae61c93
 
 
 
 
 
 
 
 
fb8cc1d
 
 
ae61c93
ed9d237
fb8cc1d
 
ae61c93
fb8cc1d
 
 
ae61c93
ed9d237
fb8cc1d
4c022fe
 
 
 
 
 
 
 
ed9d237
4c022fe
 
 
 
 
 
 
 
3c42b4c
4c022fe
 
 
 
 
 
 
 
 
 
ed9d237
4c022fe
 
 
7b1dafa
4c022fe
 
7b1dafa
3c42b4c
4c022fe
 
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
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
import math
import random
import os
import json
import time
import argparse
import torch
import numpy as np
from torchvision import transforms

from models.region_diffusion import RegionDiffusion
from utils.attention_utils import get_token_maps
from utils.richtext_utils import seed_everything, parse_json, get_region_diffusion_input,\
    get_attention_control_input, get_gradient_guidance_input


import gradio as gr
from PIL import Image, ImageOps


help_text = """
If you are encountering an error or not achieving your desired outcome, here are some potential reasons and recommendations to consider:
1. If you format only a portion of a word rather than the complete word, an error may occur. 
2. The token map may not always accurately capture the region of the formatted tokens. If you're experiencing this problem, experiment with selecting more or fewer tokens to expand or reduce the area covered by the token maps.
3. If you use font color and get completely corrupted results, you may consider decrease the color weight lambda.
4. Consider using a different seed.
"""


def main():
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model = RegionDiffusion(device)

    def generate(
        text_input: str,
        negative_text: str,
        height: int,
        width: int,
        seed: int,
        steps: int,
        guidance_weight: float,
        color_guidance_weight: float,
    ):
        run_dir = 'results/'
        # Load region diffusion model.
        steps = 41 if not steps else steps
        guidance_weight = 8.5 if not guidance_weight else guidance_weight

        # parse json to span attributes
        base_text_prompt, style_text_prompts, footnote_text_prompts, footnote_target_tokens,\
            color_text_prompts, color_names, color_rgbs, size_text_prompts_and_sizes, use_grad_guidance = parse_json(
                json.loads(text_input), device)

        # create control input for region diffusion
        region_text_prompts, region_target_token_ids, base_tokens = get_region_diffusion_input(
            model, base_text_prompt, style_text_prompts, footnote_text_prompts,
            footnote_target_tokens, color_text_prompts, color_names)

        # create control input for cross attention
        text_format_dict = get_attention_control_input(
            model, base_tokens, size_text_prompts_and_sizes)

        # create control input for region guidance
        text_format_dict, color_target_token_ids = get_gradient_guidance_input(
            model, base_tokens, color_text_prompts, color_rgbs, text_format_dict, color_guidance_weight=color_guidance_weight)

        seed_everything(seed)

        # get token maps from plain text to image generation.
        begin_time = time.time()
        if model.attention_maps is None:
            model.register_evaluation_hooks()
        else:
            model.reset_attention_maps()
        plain_img = model.produce_attn_maps([base_text_prompt], [negative_text],
                                            height=height, width=width, num_inference_steps=steps,
                                            guidance_scale=guidance_weight)
        print('time lapses to get attention maps: %.4f' % (time.time()-begin_time))
        color_obj_masks, _ = get_token_maps(
            model.attention_maps, run_dir, width//8, height//8, color_target_token_ids, seed)
        model.masks, token_maps = get_token_maps(
            model.attention_maps, run_dir, width//8, height//8, region_target_token_ids, seed, base_tokens)
        color_obj_masks = [transforms.functional.resize(color_obj_mask, (height, width),
                                                        interpolation=transforms.InterpolationMode.BICUBIC,
                                                        antialias=True)
                        for color_obj_mask in color_obj_masks]
        text_format_dict['color_obj_atten'] = color_obj_masks
        model.remove_evaluation_hooks()

        # generate image from rich text
        begin_time = time.time()
        seed_everything(seed)
        rich_img = model.prompt_to_img(region_text_prompts, [negative_text],
                                    height=height, width=width, num_inference_steps=steps,
                                    guidance_scale=guidance_weight, use_grad_guidance=use_grad_guidance,
                                    text_format_dict=text_format_dict)
        print('time lapses to generate image from rich text: %.4f' %
            (time.time()-begin_time))
        cat_img = np.concatenate([plain_img[0], rich_img[0]], 1)
        return [cat_img, token_maps]

    with gr.Blocks() as demo:
        gr.HTML("""<h1 style="font-weight: 900; margin-bottom: 7px;">Expressive Text-to-Image Generation with Rich Text</h1>
                   <p> Visit our <a href="https://rich-text-to-image.github.io/rich-text-to-json.html">rich-text-to-json interface</a> to generate rich-text JSON input.<p/>
                    <p> Project webpage: https://rich-text-to-image.github.io/<p/>""")
        with gr.Row():
            with gr.Column():
                text_input = gr.Textbox(
                    label='Rich-text JSON Input',
                    max_lines=1,
                    placeholder='Example: \'{"ops":[{"insert":"a Gothic "},{"attributes":{"color":"#b26b00"},"insert":"church"},{"insert":" in a the sunset with a beautiful landscape in the background.\n"}]}\'')
                negative_prompt = gr.Textbox(
                    label='Negative Prompt',
                    max_lines=1,
                    placeholder='')
                seed = gr.Slider(label='Seed',
                                 minimum=0,
                                 maximum=100000,
                                 step=1,
                                 value=6)
                color_guidance_weight = gr.Slider(label='Color weight lambda',
                                                  minimum=0,
                                                  maximum=2,
                                                  step=0.1,
                                                  value=0.5)
                with gr.Accordion('Other Parameters', open=False):
                    steps = gr.Slider(label='Number of Steps',
                                          minimum=0,
                                          maximum=500,
                                          step=1,
                                          value=41)
                    guidance_weight = gr.Slider(label='CFG weight',
                                               minimum=0,
                                               maximum=50,
                                               step=0.1,
                                               value=8.5)
                    width = gr.Dropdown(choices=[512, 768, 896],
                                    value=512,
                                    label='Width',
                                    visible=True)
                    height = gr.Dropdown(choices=[512, 768, 896],
                                    value=512,
                                    label='height',
                                    visible=True)
                    
                with gr.Row():
                    with gr.Column(scale=1, min_width=100):
                        generate_button = gr.Button("Generate")

            with gr.Column():
                result = gr.Image(label='Result')
                token_map = gr.Image(label='TokenMap')

        with gr.Row():
            gr.Markdown(help_text)

        with gr.Row():
            examples = [
                [
                    '{"ops":[{"insert":"a "},{"attributes":{"font":"slabo"},"insert":"night sky filled with stars"},{"insert":" above a "},{"attributes":{"font":"roboto"},"insert":"turbulent sea with giant waves"}]}',
                    '',
                    512,
                    512,
                    6,
                    1,
                ],
                [
                    '{"ops":[{"attributes":{"link":"the awe-inspiring sky and ocean in the style of J.M.W. Turner"},"insert":"the awe-inspiring sky and sea"},{"insert":" by "},{"attributes":{"font":"mirza"},"insert":"a coast with flowers and grasses in spring"}]}',
                    '',
                    512,
                    512,
                    9,
                    1,
                ],
                [
                    '{"ops":[{"insert":"a Gothic "},{"attributes":{"color":"#b26b00"},"insert":"church"},{"insert":" in a the sunset with a beautiful landscape in the background."}]}',
                    '',
                    512,
                    512,
                    6,
                    1,
                ],
                [
                    '{"ops": [{"insert": "A pizza with "}, {"attributes": {"size": "50px"}, "insert": "pineapples"}, {"insert": ", pepperonis, and mushrooms on the top, 4k, photorealistic"}]}',
                    'blurry, art, painting, rendering, drawing, sketch, ugly, duplicate, morbid, mutilated, mutated, deformed, disfigured low quality, worst quality',
                    768,
                    896,
                    6,
                    1,
                ],
                [
                    '{"ops":[{"insert":"a "},{"attributes":{"font":"mirza"},"insert":"beautiful garden"},{"insert":" with a "},{"attributes":{"font":"roboto"},"insert":"snow mountain in the background"},{"insert":""}]}',
                    '',
                    512,
                    512,
                    3,
                    1,
                ],
                [
                    '{"ops":[{"insert":"A close-up 4k dslr photo of a "},{"attributes":{"link":"A cat wearing sunglasses and a bandana around its neck."},"insert":"cat"},{"insert":" riding a scooter. Palm trees in the background."}]}',
                    '',
                    512,
                    512,
                    6,
                    1,
                ],
            ]
            gr.Examples(examples=examples,
                        inputs=[
                            text_input,
                            negative_prompt,
                            height,
                            width,
                            seed,
                            color_guidance_weight,
                        ],
                        outputs=[
                            result,
                            token_map,
                        ],
                        fn=generate,
                        # cache_examples=True,
                        examples_per_page=20)

        generate_button.click(
            fn=generate,
            inputs=[
                text_input,
                negative_prompt,
                height,
                width,
                seed,
                steps,
                guidance_weight,
                color_guidance_weight,
            ],
            outputs=[result, token_map],
        )

    demo.queue(concurrency_count=1)
    demo.launch(share=False)


if __name__ == "__main__":
    main()