File size: 5,643 Bytes
617065a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import gc
import os
import re
import shutil

import gradio as gr
import requests
import torch

from dreamcreature.pipeline import create_args, load_pipeline


def download_file(url, local_path):
    if os.path.exists(local_path):
        return

    with requests.get(url, stream=True) as r:
        with open(local_path, 'wb') as f:
            shutil.copyfileobj(r.raw, f)

    # Example usage


parser = argparse.ArgumentParser()
parser.add_argument('--model_name', default='dreamcreature-sd1.5-cub200')
parser.add_argument('--checkpoint', default='checkpoint-74900')
opt = parser.parse_args()

model_name = opt.model_name
checkpoint_name = opt.checkpoint

repo_url = f"https://huggingface.co/kamwoh/{model_name}/resolve/main"
file_url = repo_url + f"/{checkpoint_name}/pytorch_model.bin"
local_path = f"{model_name}/{checkpoint_name}/pytorch_model.bin"
os.makedirs(f"{model_name}/{checkpoint_name}", exist_ok=True)
download_file(file_url, local_path)

file_url = repo_url + f"/{checkpoint_name}/pytorch_model_1.bin"
local_path = f"{model_name}/{checkpoint_name}/pytorch_model_1.bin"
download_file(file_url, local_path)

OUTPUT_DIR = model_name

args = create_args(OUTPUT_DIR)
if 'dpo' in OUTPUT_DIR:
    args.unet_path = "mhdang/dpo-sd1.5-text2image-v1"

pipe = load_pipeline(args, torch.float16, 'cuda')
pipe = pipe.to(torch.float16)

pipe.verbose = True
pipe.v = 're'
pipe.num_k_per_part = 200

MAPPING = {
    'body': 0,
    'tail': 1,
    'head': 2,
    'wing': 4,
    'leg': 6
}

ID2NAME = open('data/cub200_2011/class_names.txt').readlines()
ID2NAME = [line.strip() for line in ID2NAME]


def process_text(text):
    pattern = r"<([^:>]+):(\d+)>"
    result = text
    offset = 0

    part2id = []

    for match in re.finditer(pattern, text):
        key = match.group(1)
        clsid = int(match.group(2))
        clsid = min(max(clsid, 1), 200)  # must be 1~200

        replacement = f"<{MAPPING[key]}:{clsid - 1}>"
        start, end = match.span()

        # Adjust the start and end positions based on the offset from previous replacements
        start += offset
        end += offset

        # Replace the matched text with the replacement
        result = result[:start] + replacement + result[end:]

        # Update the offset for the next replacement
        offset += len(replacement) - (end - start)

        part2id.append(f'{key}: {ID2NAME[clsid - 1]}')

    return result, part2id


def generate_images(prompt, negative_prompt, num_inference_steps, guidance_scale, num_images, seed):
    generator = torch.Generator(device='cuda')
    generator = generator.manual_seed(int(seed))

    try:
        prompt, part2id = process_text(prompt)
        negative_prompt, _ = process_text(negative_prompt)

        images = pipe(prompt,
                      negative_prompt=negative_prompt, generator=generator,
                      num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
                      num_images_per_prompt=num_images).images
    except Exception as e:
        raise gr.Error(f"Probably due to the prompt have invalid input, please follow the instruction. "
                       f"The error message: {e}")
    finally:
        gc.collect()
        torch.cuda.empty_cache()

    return images, '; '.join(part2id)


with gr.Blocks(title="DreamCreature") as demo:
    with gr.Row():
        gr.Markdown(
            """
            # DreamCreature (CUB-200-2011)
            To create your own creature, you can type:

            `"a photo of a <head:id> <wing:id> bird"` where `id` ranges from 1~200 (200 classes corresponding to CUB-200-2011)

            For instance `"a photo of a <head:17> <wing:18> bird"` using head of `cardinal (17)` and wing of `spotted catbird (18)`

            Please see `id` in https://github.com/kamwoh/dreamcreature/blob/master/src/data/cub200_2011/class_names.txt

            You can also try any prompt you like such as:

            Sub-concept transfer: `"a photo of a <wing:17> cat"`

            Inspiring design: `"a photo of a <head:101> <wing:191> teddy bear"`

            (Experimental) You can also use two parts together such as:

            `"a photo of a <head:17> <head:18> bird"` mixing head of `cardinal (17)` and `spotted catbird (18)`

            The current available parts are: `head`, `body`, `wing`, `tail`, and `leg`

            """)
    with gr.Column():
        with gr.Row():
            with gr.Group():
                prompt = gr.Textbox(label="Prompt", value="a photo of a <head:101> <wing:191> teddy bear")
                negative_prompt = gr.Textbox(label="Negative Prompt",
                                             value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
                num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=30, label="Num Inference Steps")
                guidance_scale = gr.Slider(minimum=2, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
                num_images = gr.Slider(minimum=1, maximum=4, step=1, value=4, label="Number of Images")
                seed = gr.Number(label="Seed", value=777881414)
                button = gr.Button()

            with gr.Column():
                output_images = gr.Gallery(columns=4, label='Output')
                markdown_labels = gr.Markdown("")

    button.click(fn=generate_images,
                 inputs=[prompt, negative_prompt, num_inference_steps, guidance_scale, num_images,
                         seed], outputs=[output_images, markdown_labels], show_progress=True)

demo.queue().launch(inline=False, share=True, debug=True, server_name='0.0.0.0')