Update app.py -- first released version
Browse files
app.py
CHANGED
@@ -1,96 +1,248 @@
|
|
1 |
import gradio as gr
|
2 |
import numpy as np
|
3 |
import random
|
4 |
-
from diffusers import DiffusionPipeline
|
5 |
import torch
|
6 |
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
pipe.enable_xformers_memory_efficient_attention()
|
13 |
-
pipe = pipe.to(device)
|
14 |
-
else:
|
15 |
-
pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", use_safetensors=True)
|
16 |
-
pipe = pipe.to(device)
|
17 |
|
18 |
-
|
19 |
-
|
20 |
|
21 |
-
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
if randomize_seed:
|
24 |
seed = random.randint(0, MAX_SEED)
|
25 |
-
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
-
return image
|
39 |
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
|
|
44 |
]
|
45 |
|
|
|
|
|
|
|
46 |
css="""
|
47 |
#col-container {
|
48 |
margin: 0 auto;
|
49 |
-
max-width:
|
50 |
}
|
51 |
"""
|
52 |
|
53 |
-
if torch.cuda.is_available():
|
54 |
-
power_device = "GPU"
|
55 |
-
else:
|
56 |
-
power_device = "CPU"
|
57 |
-
|
58 |
with gr.Blocks(css=css) as demo:
|
59 |
|
60 |
with gr.Column(elem_id="col-container"):
|
61 |
gr.Markdown(f"""
|
62 |
-
#
|
63 |
Currently running on {power_device}.
|
64 |
""")
|
65 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
with gr.Row():
|
67 |
-
|
68 |
-
with gr.Column():
|
69 |
-
prompt = gr.Text(
|
70 |
-
label="Prompt",
|
71 |
-
show_label=False,
|
72 |
-
max_lines=1,
|
73 |
-
placeholder="Enter your prompt",
|
74 |
-
container=False,
|
75 |
-
)
|
76 |
-
prompt2 = gr.Text(
|
77 |
-
label="Prompt2",
|
78 |
-
show_label=False,
|
79 |
-
max_lines=1,
|
80 |
-
placeholder="Enter your prompt for right character",
|
81 |
-
container=False,
|
82 |
-
)
|
83 |
|
84 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
|
86 |
result = gr.Image(label="Result", show_label=False)
|
87 |
|
88 |
with gr.Accordion("Advanced Settings", open=False):
|
89 |
-
|
90 |
negative_prompt = gr.Text(
|
91 |
-
label="Negative prompt",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
92 |
max_lines=1,
|
93 |
-
|
94 |
visible=False,
|
95 |
)
|
96 |
|
@@ -106,49 +258,60 @@ with gr.Blocks(css=css) as demo:
|
|
106 |
|
107 |
with gr.Row():
|
108 |
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
)
|
116 |
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
)
|
124 |
|
125 |
-
|
126 |
-
|
127 |
-
guidance_scale = gr.Slider(
|
128 |
-
label="Guidance scale",
|
129 |
-
minimum=0.0,
|
130 |
-
maximum=10.0,
|
131 |
-
step=0.1,
|
132 |
-
value=0.0,
|
133 |
-
)
|
134 |
-
|
135 |
-
num_inference_steps = gr.Slider(
|
136 |
-
label="Number of inference steps",
|
137 |
-
minimum=1,
|
138 |
-
maximum=12,
|
139 |
-
step=1,
|
140 |
-
value=2,
|
141 |
-
)
|
142 |
-
|
143 |
gr.Examples(
|
144 |
-
|
|
|
145 |
inputs = [prompt]
|
146 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
147 |
|
148 |
run_button.click(
|
149 |
-
fn =
|
150 |
-
inputs = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
outputs = [result]
|
152 |
)
|
153 |
|
154 |
-
demo.queue().launch()
|
|
|
1 |
import gradio as gr
|
2 |
import numpy as np
|
3 |
import random
|
|
|
4 |
import torch
|
5 |
|
6 |
+
import io, json
|
7 |
+
from PIL import Image
|
8 |
+
import os.path
|
9 |
+
from weight_fusion import compose_concepts
|
10 |
+
from regionally_controlable_sampling import sample_image, build_model, prepare_text
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
13 |
+
power_device = "GPU" if torch.cuda.is_available() else "CPU"
|
14 |
|
15 |
+
MAX_SEED = 100_000
|
16 |
|
17 |
+
def generate(region1_concept,
|
18 |
+
region2_concept,
|
19 |
+
prompt,
|
20 |
+
region1_prompt,
|
21 |
+
region2_prompt,
|
22 |
+
negative_prompt,
|
23 |
+
region_neg_prompt,
|
24 |
+
seed,
|
25 |
+
randomize_seed,
|
26 |
+
sketch_adaptor_weight,
|
27 |
+
keypose_adaptor_weight
|
28 |
+
):
|
29 |
+
|
30 |
if randomize_seed:
|
31 |
seed = random.randint(0, MAX_SEED)
|
32 |
+
|
33 |
+
region1_concept, region2_concept = region1_concept.lower(), region2_concept.lower()
|
34 |
+
pretrained_model = merge(region1_concept, region2_concept)
|
35 |
+
|
36 |
+
keypose_condition = 'multi-concept/pose_data/two_apart.png'
|
37 |
+
region1 = '[0, 0, 512, 290]'
|
38 |
+
region2 = '[0, 650, 512, 910]'
|
39 |
+
|
40 |
+
region1_prompt = f'[<{region1_concept}1> <{region1_concept}2>, {region1_prompt}]'
|
41 |
+
region2_prompt = f'[<{region2_concept}1> <{region2_concept}2>, {region2_prompt}]'
|
42 |
+
prompt_rewrite=f"{region1_prompt}-*-{region_neg_prompt}-*-{region1}|{region2_prompt}-*-{region_neg_prompt}-*-{region2}"
|
43 |
+
|
44 |
+
result = infer(pretrained_model,
|
45 |
+
prompt,
|
46 |
+
prompt_rewrite,
|
47 |
+
negative_prompt,
|
48 |
+
seed,
|
49 |
+
keypose_condition,
|
50 |
+
keypose_adaptor_weight,
|
51 |
+
# sketch_condition,
|
52 |
+
# sketch_adaptor_weight,
|
53 |
+
)
|
54 |
|
55 |
+
return result
|
56 |
+
|
57 |
+
def merge(concept1, concept2):
|
58 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
59 |
+
c1, c2 = sorted([concept1, concept2])
|
60 |
+
assert c1!=c2
|
61 |
+
merge_name = c1+'_'+c2
|
62 |
+
|
63 |
+
save_path = f'experiments/multi-concept/{merge_name}'
|
64 |
+
|
65 |
+
if os.path.isdir(save_path):
|
66 |
+
print(f'{save_path} already exists. Collecting merged weights from existing weights...')
|
67 |
+
|
68 |
+
else:
|
69 |
+
os.makedirs(save_path)
|
70 |
+
json_path = os.path.join(save_path,'merge_config.json')
|
71 |
+
alpha = 1.8
|
72 |
+
data = [
|
73 |
+
{
|
74 |
+
"lora_path": f"experiments/single-concept/{c1}/models/edlora_model-latest.pth",
|
75 |
+
"unet_alpha": alpha,
|
76 |
+
"text_encoder_alpha": alpha,
|
77 |
+
"concept_name": f"<{c1}1> <{c1}2>"
|
78 |
+
},
|
79 |
+
{
|
80 |
+
"lora_path": f"experiments/single-concept/{c2}/models/edlora_model-latest.pth",
|
81 |
+
"unet_alpha": alpha,
|
82 |
+
"text_encoder_alpha": alpha,
|
83 |
+
"concept_name": f"<{c2}1> <{c2}2>"
|
84 |
+
}
|
85 |
+
]
|
86 |
+
with io.open(json_path,'w',encoding='utf8') as outfile:
|
87 |
+
json.dump(data, outfile, indent = 4, ensure_ascii=False)
|
88 |
+
|
89 |
+
compose_concepts(
|
90 |
+
concept_cfg=json_path,
|
91 |
+
optimize_textenc_iters=500,
|
92 |
+
optimize_unet_iters=50,
|
93 |
+
pretrained_model_path="nitrosocke/mo-di-diffusion",
|
94 |
+
save_path=save_path,
|
95 |
+
suffix='base',
|
96 |
+
device=device,
|
97 |
+
)
|
98 |
+
print(f'Merged weight for {c1}+{c2} saved in {save_path}!\n\n')
|
99 |
+
|
100 |
+
modelbase_path = os.path.join(save_path,'combined_model_base')
|
101 |
+
assert os.path.isdir(modelbase_path)
|
102 |
+
|
103 |
+
# save_path = 'experiments/multi-concept/elsa_moana_weight18/combined_model_base'
|
104 |
+
return modelbase_path
|
105 |
+
|
106 |
+
def infer(pretrained_model,
|
107 |
+
prompt,
|
108 |
+
prompt_rewrite,
|
109 |
+
negative_prompt='',
|
110 |
+
seed=16141,
|
111 |
+
keypose_condition=None,
|
112 |
+
keypose_adaptor_weight=1.0,
|
113 |
+
sketch_condition=None,
|
114 |
+
sketch_adaptor_weight=0.0,
|
115 |
+
region_sketch_adaptor_weight='',
|
116 |
+
region_keypose_adaptor_weight=''
|
117 |
+
):
|
118 |
+
|
119 |
+
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
|
120 |
+
pipe = build_model(pretrained_model, device)
|
121 |
+
|
122 |
+
if sketch_condition is not None and os.path.exists(sketch_condition):
|
123 |
+
sketch_condition = Image.open(sketch_condition).convert('L')
|
124 |
+
width_sketch, height_sketch = sketch_condition.size
|
125 |
+
print('use sketch condition')
|
126 |
+
else:
|
127 |
+
sketch_condition, width_sketch, height_sketch = None, 0, 0
|
128 |
+
print('skip sketch condition')
|
129 |
+
|
130 |
+
if keypose_condition is not None and os.path.exists(keypose_condition):
|
131 |
+
keypose_condition = Image.open(keypose_condition).convert('RGB')
|
132 |
+
width_pose, height_pose = keypose_condition.size
|
133 |
+
print('use pose condition')
|
134 |
+
else:
|
135 |
+
keypose_condition, width_pose, height_pose = None, 0, 0
|
136 |
+
print('skip pose condition')
|
137 |
+
|
138 |
+
if width_sketch != 0 and width_pose != 0:
|
139 |
+
assert width_sketch == width_pose and height_sketch == height_pose, 'conditions should be same size'
|
140 |
+
width, height = max(width_pose, width_sketch), max(height_pose, height_sketch)
|
141 |
+
kwargs = {
|
142 |
+
'sketch_condition': sketch_condition,
|
143 |
+
'keypose_condition': keypose_condition,
|
144 |
+
'height': height,
|
145 |
+
'width': width,
|
146 |
+
}
|
147 |
+
|
148 |
+
prompts = [prompt]
|
149 |
+
prompts_rewrite = [prompt_rewrite]
|
150 |
+
input_prompt = [prepare_text(p, p_w, height, width) for p, p_w in zip(prompts, prompts_rewrite)]
|
151 |
+
save_prompt = input_prompt[0][0]
|
152 |
+
print(save_prompt)
|
153 |
+
|
154 |
+
image = sample_image(
|
155 |
+
pipe,
|
156 |
+
input_prompt=input_prompt,
|
157 |
+
input_neg_prompt=[negative_prompt] * len(input_prompt),
|
158 |
+
generator=torch.Generator(device).manual_seed(seed),
|
159 |
+
sketch_adaptor_weight=sketch_adaptor_weight,
|
160 |
+
region_sketch_adaptor_weight=region_sketch_adaptor_weight,
|
161 |
+
keypose_adaptor_weight=keypose_adaptor_weight,
|
162 |
+
region_keypose_adaptor_weight=region_keypose_adaptor_weight,
|
163 |
+
**kwargs)
|
164 |
|
165 |
+
return image[0]
|
166 |
|
167 |
+
examples_context = [
|
168 |
+
'walking at Stanford university campus',
|
169 |
+
'in a castle',
|
170 |
+
'in the forest',
|
171 |
+
'in front of Eiffel tower'
|
172 |
]
|
173 |
|
174 |
+
examples_region1 = ['wearing red hat, high resolution, best quality','bright smile, wearing pants, best quality']
|
175 |
+
examples_region2 = ['smilling, wearing blue shirt, high resolution, best quality']
|
176 |
+
|
177 |
css="""
|
178 |
#col-container {
|
179 |
margin: 0 auto;
|
180 |
+
max-width: 600px;
|
181 |
}
|
182 |
"""
|
183 |
|
|
|
|
|
|
|
|
|
|
|
184 |
with gr.Blocks(css=css) as demo:
|
185 |
|
186 |
with gr.Column(elem_id="col-container"):
|
187 |
gr.Markdown(f"""
|
188 |
+
# Orthogonal Adaptation
|
189 |
Currently running on {power_device}.
|
190 |
""")
|
191 |
+
prompt = gr.Text(
|
192 |
+
label="ContextPrompt",
|
193 |
+
show_label=False,
|
194 |
+
max_lines=1,
|
195 |
+
placeholder="Enter your context prompt for overall image",
|
196 |
+
container=False,
|
197 |
+
)
|
198 |
with gr.Row():
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
199 |
|
200 |
+
region1_concept = gr.Dropdown(
|
201 |
+
["Elsa", "Moana"],
|
202 |
+
label="Character 1",
|
203 |
+
info="Will add more characters later!"
|
204 |
+
)
|
205 |
+
region2_concept = gr.Dropdown(
|
206 |
+
["Elsa", "Moana"],
|
207 |
+
label="Character 2",
|
208 |
+
info="Will add more characters later!"
|
209 |
+
)
|
210 |
+
|
211 |
+
with gr.Row():
|
212 |
+
|
213 |
+
region1_prompt = gr.Textbox(
|
214 |
+
label="Region1 Prompt",
|
215 |
+
show_label=False,
|
216 |
+
max_lines=2,
|
217 |
+
placeholder="Enter your prompt for character 1",
|
218 |
+
container=False,
|
219 |
+
)
|
220 |
+
|
221 |
+
region2_prompt = gr.Textbox(
|
222 |
+
label="Region2 Prompt",
|
223 |
+
show_label=False,
|
224 |
+
max_lines=2,
|
225 |
+
placeholder="Enter your prompt for character 2",
|
226 |
+
container=False,
|
227 |
+
)
|
228 |
+
|
229 |
+
run_button = gr.Button("Run", scale=1)
|
230 |
|
231 |
result = gr.Image(label="Result", show_label=False)
|
232 |
|
233 |
with gr.Accordion("Advanced Settings", open=False):
|
234 |
+
|
235 |
negative_prompt = gr.Text(
|
236 |
+
label="Context Negative prompt",
|
237 |
+
max_lines=1,
|
238 |
+
value = 'saturated, cropped, worst quality, low quality',
|
239 |
+
visible=False,
|
240 |
+
)
|
241 |
+
|
242 |
+
region_neg_prompt = gr.Text(
|
243 |
+
label="Regional Negative prompt",
|
244 |
max_lines=1,
|
245 |
+
value = 'shirtless, nudity, saturated, cropped, worst quality, low quality',
|
246 |
visible=False,
|
247 |
)
|
248 |
|
|
|
258 |
|
259 |
with gr.Row():
|
260 |
|
261 |
+
sketch_adaptor_weight = gr.Slider(
|
262 |
+
label="Sketch Adapter Weight",
|
263 |
+
minimum = 0,
|
264 |
+
maximum = 1,
|
265 |
+
step=0.01,
|
266 |
+
value=0,
|
267 |
)
|
268 |
|
269 |
+
keypose_adaptor_weight = gr.Slider(
|
270 |
+
label="Keypose Adapter Weight",
|
271 |
+
minimum = 0,
|
272 |
+
maximum = 1,
|
273 |
+
step= 0.01,
|
274 |
+
value=1.0,
|
275 |
)
|
276 |
|
277 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
278 |
gr.Examples(
|
279 |
+
label = 'Context Prompt example',
|
280 |
+
examples = examples_context,
|
281 |
inputs = [prompt]
|
282 |
+
)
|
283 |
+
|
284 |
+
with gr.Row():
|
285 |
+
gr.Examples(
|
286 |
+
label = 'Region1 Prompt example',
|
287 |
+
examples = examples_region1,
|
288 |
+
inputs = [region1_prompt]
|
289 |
+
)
|
290 |
+
|
291 |
+
gr.Examples(
|
292 |
+
label = 'Region2 Prompt example',
|
293 |
+
examples = [examples_region2],
|
294 |
+
inputs = [region2_prompt]
|
295 |
+
)
|
296 |
+
|
297 |
|
298 |
run_button.click(
|
299 |
+
fn = generate,
|
300 |
+
inputs = [region1_concept,
|
301 |
+
region2_concept,
|
302 |
+
prompt,
|
303 |
+
region1_prompt,
|
304 |
+
region2_prompt,
|
305 |
+
negative_prompt,
|
306 |
+
region_neg_prompt,
|
307 |
+
seed,
|
308 |
+
randomize_seed,
|
309 |
+
# sketch_condition,
|
310 |
+
# keypose_condition,
|
311 |
+
sketch_adaptor_weight,
|
312 |
+
keypose_adaptor_weight
|
313 |
+
],
|
314 |
outputs = [result]
|
315 |
)
|
316 |
|
317 |
+
demo.queue().launch(share=True)
|