furkan gözükara ev pc
commited on
Commit
•
a2f9064
1
Parent(s):
b40e904
v1
Browse files- .gitignore +5 -0
- app.py +164 -170
.gitignore
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
venv
|
2 |
+
.git
|
3 |
+
.vs
|
4 |
+
outputs
|
5 |
+
previewer
|
app.py
CHANGED
@@ -8,29 +8,42 @@ from typing import List
|
|
8 |
from diffusers.utils import numpy_to_pil
|
9 |
from diffusers import StableCascadeDecoderPipeline, StableCascadePriorPipeline
|
10 |
from diffusers.pipelines.wuerstchen import DEFAULT_STAGE_C_TIMESTEPS
|
11 |
-
import spaces
|
12 |
from previewer.modules import Previewer
|
13 |
-
import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
-
os.environ['TOKENIZERS_PARALLELISM'] = 'false'
|
16 |
|
17 |
-
|
18 |
-
DESCRIPTION
|
19 |
if not torch.cuda.is_available():
|
20 |
-
DESCRIPTION += "
|
21 |
|
22 |
MAX_SEED = np.iinfo(np.int32).max
|
23 |
-
|
24 |
-
MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "1536"))
|
25 |
-
USE_TORCH_COMPILE = False
|
26 |
-
ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD") == "1"
|
27 |
PREVIEW_IMAGES = True
|
28 |
|
29 |
dtype = torch.bfloat16
|
30 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
31 |
if torch.cuda.is_available():
|
32 |
-
prior_pipeline = StableCascadePriorPipeline.from_pretrained("stabilityai/stable-cascade-prior", torch_dtype=dtype)
|
33 |
-
decoder_pipeline = StableCascadeDecoderPipeline.from_pretrained("stabilityai/stable-cascade", torch_dtype=dtype)
|
|
|
|
|
34 |
|
35 |
if ENABLE_CPU_OFFLOAD:
|
36 |
prior_pipeline.enable_model_cpu_offload()
|
@@ -66,7 +79,6 @@ def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
|
|
66 |
seed = random.randint(0, MAX_SEED)
|
67 |
return seed
|
68 |
|
69 |
-
@spaces.GPU
|
70 |
def generate(
|
71 |
prompt: str,
|
72 |
negative_prompt: str = "",
|
@@ -74,172 +86,159 @@ def generate(
|
|
74 |
width: int = 1024,
|
75 |
height: int = 1024,
|
76 |
prior_num_inference_steps: int = 30,
|
77 |
-
# prior_timesteps: List[float] = None,
|
78 |
prior_guidance_scale: float = 4.0,
|
79 |
decoder_num_inference_steps: int = 12,
|
80 |
-
# decoder_timesteps: List[float] = None,
|
81 |
decoder_guidance_scale: float = 0.0,
|
82 |
-
|
83 |
-
|
84 |
-
) -> PIL.Image.Image:
|
85 |
-
#
|
86 |
-
#
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
for _ in range(len(DEFAULT_STAGE_C_TIMESTEPS)):
|
105 |
-
r = next(prior_output)
|
106 |
-
if isinstance(r, list):
|
107 |
-
yield r[0]
|
108 |
-
prior_output = r
|
109 |
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
# timesteps=decoder_timesteps,
|
115 |
-
guidance_scale=decoder_guidance_scale,
|
116 |
-
negative_prompt=negative_prompt,
|
117 |
-
generator=generator,
|
118 |
-
output_type="pil",
|
119 |
-
).images
|
120 |
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
"width": width,
|
131 |
-
"height": height,
|
132 |
-
"prior_guidance_scale": prior_guidance_scale,
|
133 |
-
"decoder_num_inference_steps": decoder_num_inference_steps,
|
134 |
-
"decoder_guidance_scale": decoder_guidance_scale,
|
135 |
-
"num_images_per_prompt": num_images_per_prompt,
|
136 |
-
},
|
137 |
-
)
|
138 |
|
139 |
-
|
|
|
140 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
141 |
|
142 |
-
|
143 |
-
|
144 |
-
"An astronaut riding a green horse",
|
145 |
-
"A mecha robot in a favela by Tarsila do Amaral",
|
146 |
-
"The sprirt of a Tamagotchi wandering in the city of Los Angeles",
|
147 |
-
"A delicious feijoada ramen dish"
|
148 |
-
]
|
149 |
|
150 |
-
with gr.Blocks() as
|
151 |
-
gr.
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
visible=os.getenv("SHOW_DUPLICATE_BUTTON") == "1",
|
156 |
-
)
|
157 |
-
with gr.Group():
|
158 |
-
with gr.Row():
|
159 |
prompt = gr.Text(
|
160 |
label="Prompt",
|
161 |
-
show_label=False,
|
162 |
-
max_lines=1,
|
163 |
placeholder="Enter your prompt",
|
164 |
-
container=False,
|
165 |
-
)
|
166 |
-
run_button = gr.Button("Run", scale=0)
|
167 |
-
result = gr.Image(label="Result", show_label=False)
|
168 |
-
with gr.Accordion("Advanced options", open=False):
|
169 |
-
negative_prompt = gr.Text(
|
170 |
-
label="Negative prompt",
|
171 |
-
max_lines=1,
|
172 |
-
placeholder="Enter a Negative Prompt",
|
173 |
-
)
|
174 |
-
|
175 |
-
seed = gr.Slider(
|
176 |
-
label="Seed",
|
177 |
-
minimum=0,
|
178 |
-
maximum=MAX_SEED,
|
179 |
-
step=1,
|
180 |
-
value=0,
|
181 |
-
)
|
182 |
-
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
183 |
-
with gr.Row():
|
184 |
-
width = gr.Slider(
|
185 |
-
label="Width",
|
186 |
-
minimum=1024,
|
187 |
-
maximum=MAX_IMAGE_SIZE,
|
188 |
-
step=512,
|
189 |
-
value=1024,
|
190 |
-
)
|
191 |
-
height = gr.Slider(
|
192 |
-
label="Height",
|
193 |
-
minimum=1024,
|
194 |
-
maximum=MAX_IMAGE_SIZE,
|
195 |
-
step=512,
|
196 |
-
value=1024,
|
197 |
-
)
|
198 |
-
num_images_per_prompt = gr.Slider(
|
199 |
-
label="Number of Images",
|
200 |
-
minimum=1,
|
201 |
-
maximum=2,
|
202 |
-
step=1,
|
203 |
-
value=1,
|
204 |
)
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
value=4.0,
|
212 |
-
)
|
213 |
-
prior_num_inference_steps = gr.Slider(
|
214 |
-
label="Prior Inference Steps",
|
215 |
-
minimum=10,
|
216 |
-
maximum=30,
|
217 |
-
step=1,
|
218 |
-
value=20,
|
219 |
)
|
220 |
|
221 |
-
|
222 |
-
label="
|
223 |
minimum=0,
|
224 |
-
maximum=
|
225 |
-
step=0.1,
|
226 |
-
value=0.0,
|
227 |
-
)
|
228 |
-
decoder_num_inference_steps = gr.Slider(
|
229 |
-
label="Decoder Inference Steps",
|
230 |
-
minimum=4,
|
231 |
-
maximum=12,
|
232 |
step=1,
|
233 |
-
value=
|
234 |
)
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
243 |
|
244 |
inputs = [
|
245 |
prompt,
|
@@ -253,7 +252,8 @@ with gr.Blocks() as demo:
|
|
253 |
decoder_num_inference_steps,
|
254 |
# decoder_timesteps,
|
255 |
decoder_guidance_scale,
|
256 |
-
|
|
|
257 |
]
|
258 |
gr.on(
|
259 |
triggers=[prompt.submit, negative_prompt.submit, run_button.click],
|
@@ -268,12 +268,6 @@ with gr.Blocks() as demo:
|
|
268 |
outputs=result,
|
269 |
api_name="run",
|
270 |
)
|
271 |
-
|
272 |
-
with gr.Blocks(css="style.css") as demo_with_history:
|
273 |
-
with gr.Tab("App"):
|
274 |
-
demo.render()
|
275 |
-
with gr.Tab("Past generations"):
|
276 |
-
user_history.render()
|
277 |
-
|
278 |
if __name__ == "__main__":
|
279 |
-
|
|
|
8 |
from diffusers.utils import numpy_to_pil
|
9 |
from diffusers import StableCascadeDecoderPipeline, StableCascadePriorPipeline
|
10 |
from diffusers.pipelines.wuerstchen import DEFAULT_STAGE_C_TIMESTEPS
|
|
|
11 |
from previewer.modules import Previewer
|
12 |
+
import os
|
13 |
+
import datetime
|
14 |
+
import json
|
15 |
+
import io
|
16 |
+
import argparse # Import the argparse library
|
17 |
+
|
18 |
+
# Set up argument parser
|
19 |
+
parser = argparse.ArgumentParser(description="Gradio interface for text-to-image generation with optional features.")
|
20 |
+
parser.add_argument("--share", action="store_true", help="Enable Gradio sharing.")
|
21 |
+
parser.add_argument("--lowvram", action="store_true", help="Enable CPU offload for model operations.")
|
22 |
+
parser.add_argument("--torch_compile", action="store_true", help="Enable CPU offload for model operations.")
|
23 |
+
|
24 |
+
# Parse arguments
|
25 |
+
args = parser.parse_args()
|
26 |
+
share = args.share
|
27 |
+
ENABLE_CPU_OFFLOAD = args.lowvram # Use the offload argument to toggle ENABLE_CPU_OFFLOAD
|
28 |
+
USE_TORCH_COMPILE = args.torch_compile # Use the offload argument to toggle ENABLE_CPU_OFFLOAD
|
29 |
|
|
|
30 |
|
31 |
+
os.environ['TOKENIZERS_PARALLELISM'] = 'false'
|
32 |
+
DESCRIPTION = "<p style=\"font-size:14px\">Stable Cascade Modified By SECourses - Unofficial demo for <a href='https://huggingface.co/stabilityai/stable-cascade' target='_blank'>Stable Casacade</a>, a new high resolution text-to-image model by Stability AI, built on the Würstchen architecture.<br/> Some tips: Higher batch size working great with fast speed and not much VRAM usage - Not all resolutions working e.g. 1920x1080 fails but 1920x1152 works<br/>Supports high resolutions very well such as 1536x1536</p>"
|
33 |
if not torch.cuda.is_available():
|
34 |
+
DESCRIPTION += "<br/><p>Running on CPU 🥶</p>"
|
35 |
|
36 |
MAX_SEED = np.iinfo(np.int32).max
|
37 |
+
MAX_IMAGE_SIZE = 2048
|
|
|
|
|
|
|
38 |
PREVIEW_IMAGES = True
|
39 |
|
40 |
dtype = torch.bfloat16
|
41 |
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
42 |
if torch.cuda.is_available():
|
43 |
+
prior_pipeline = StableCascadePriorPipeline.from_pretrained("stabilityai/stable-cascade-prior", torch_dtype=dtype)
|
44 |
+
decoder_pipeline = StableCascadeDecoderPipeline.from_pretrained("stabilityai/stable-cascade", torch_dtype=dtype)
|
45 |
+
prior_pipeline.enable_xformers_memory_efficient_attention()
|
46 |
+
decoder_pipeline.enable_xformers_memory_efficient_attention()
|
47 |
|
48 |
if ENABLE_CPU_OFFLOAD:
|
49 |
prior_pipeline.enable_model_cpu_offload()
|
|
|
79 |
seed = random.randint(0, MAX_SEED)
|
80 |
return seed
|
81 |
|
|
|
82 |
def generate(
|
83 |
prompt: str,
|
84 |
negative_prompt: str = "",
|
|
|
86 |
width: int = 1024,
|
87 |
height: int = 1024,
|
88 |
prior_num_inference_steps: int = 30,
|
|
|
89 |
prior_guidance_scale: float = 4.0,
|
90 |
decoder_num_inference_steps: int = 12,
|
|
|
91 |
decoder_guidance_scale: float = 0.0,
|
92 |
+
batch_size_per_prompt: int = 2,
|
93 |
+
number_of_images_per_prompt: int = 1, # New parameter
|
94 |
+
) -> List[PIL.Image.Image]:
|
95 |
+
images = [] # Initialize an empty list to collect generated images
|
96 |
+
original_seed = seed # Store the original seed value
|
97 |
+
for i in range(number_of_images_per_prompt):
|
98 |
+
if i > 0: # Update seed for subsequent iterations
|
99 |
+
seed = random.randint(0, MAX_SEED)
|
100 |
+
generator = torch.Generator().manual_seed(seed)
|
101 |
+
prior_output = prior_pipeline(
|
102 |
+
prompt=prompt,
|
103 |
+
height=height,
|
104 |
+
width=width,
|
105 |
+
num_inference_steps=prior_num_inference_steps,
|
106 |
+
timesteps=DEFAULT_STAGE_C_TIMESTEPS,
|
107 |
+
negative_prompt=negative_prompt,
|
108 |
+
guidance_scale=prior_guidance_scale,
|
109 |
+
num_images_per_prompt=batch_size_per_prompt,
|
110 |
+
generator=generator,
|
111 |
+
callback=callback_prior,
|
112 |
+
callback_steps=callback_steps
|
113 |
+
)
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
+
if PREVIEW_IMAGES:
|
116 |
+
for _ in range(len(DEFAULT_STAGE_C_TIMESTEPS)):
|
117 |
+
r = next(prior_output)
|
118 |
+
prior_output = r
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
|
120 |
+
decoder_output = decoder_pipeline(
|
121 |
+
image_embeddings=prior_output.image_embeddings,
|
122 |
+
prompt=prompt,
|
123 |
+
num_inference_steps= decoder_num_inference_steps,
|
124 |
+
guidance_scale=decoder_guidance_scale,
|
125 |
+
negative_prompt=negative_prompt,
|
126 |
+
generator=generator,
|
127 |
+
output_type="pil",
|
128 |
+
).images
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
|
130 |
+
# Append generated images to the images list
|
131 |
+
images.extend(decoder_output)
|
132 |
|
133 |
+
# Optionally, save each image
|
134 |
+
output_folder = 'outputs'
|
135 |
+
if not os.path.exists(output_folder):
|
136 |
+
os.makedirs(output_folder)
|
137 |
+
for image in decoder_output:
|
138 |
+
# Generate timestamped filename
|
139 |
+
timestamp = datetime.datetime.now().strftime('%Y_%m_%d_%H_%M_%S_%f')
|
140 |
+
image_filename = f"{output_folder}/{timestamp}.png"
|
141 |
+
image.save(image_filename)
|
142 |
|
143 |
+
# Return the list of generated images
|
144 |
+
return images
|
|
|
|
|
|
|
|
|
|
|
145 |
|
146 |
+
with gr.Blocks() as app:
|
147 |
+
with gr.Row():
|
148 |
+
gr.Markdown(DESCRIPTION)
|
149 |
+
with gr.Row():
|
150 |
+
with gr.Column():
|
|
|
|
|
|
|
|
|
151 |
prompt = gr.Text(
|
152 |
label="Prompt",
|
|
|
|
|
153 |
placeholder="Enter your prompt",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
)
|
155 |
+
run_button = gr.Button("Generate")
|
156 |
+
|
157 |
+
# Advanced options now directly visible
|
158 |
+
negative_prompt = gr.Text(
|
159 |
+
label="Negative prompt",
|
160 |
+
placeholder="Enter a Negative Prompt",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
161 |
)
|
162 |
|
163 |
+
seed = gr.Slider(
|
164 |
+
label="Seed",
|
165 |
minimum=0,
|
166 |
+
maximum=MAX_SEED,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
167 |
step=1,
|
168 |
+
value=0,
|
169 |
)
|
170 |
+
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
171 |
+
with gr.Row():
|
172 |
+
with gr.Column():
|
173 |
+
width = gr.Slider(
|
174 |
+
label="Width",
|
175 |
+
minimum=512,
|
176 |
+
maximum=MAX_IMAGE_SIZE,
|
177 |
+
step=64,
|
178 |
+
value=1024,
|
179 |
+
)
|
180 |
+
with gr.Column():
|
181 |
+
height = gr.Slider(
|
182 |
+
label="Height",
|
183 |
+
minimum=512,
|
184 |
+
maximum=MAX_IMAGE_SIZE,
|
185 |
+
step=64,
|
186 |
+
value=1024,
|
187 |
+
)
|
188 |
+
with gr.Row():
|
189 |
+
with gr.Column():
|
190 |
+
batch_size_per_prompt = gr.Slider(
|
191 |
+
label="Batch Size",
|
192 |
+
minimum=1,
|
193 |
+
maximum=20,
|
194 |
+
step=1,
|
195 |
+
value=1,
|
196 |
+
)
|
197 |
+
with gr.Column():
|
198 |
+
number_of_images_per_prompt = gr.Slider(
|
199 |
+
label="Number Of Images To Generate",
|
200 |
+
minimum=1,
|
201 |
+
maximum=9999999,
|
202 |
+
step=1,
|
203 |
+
value=1,
|
204 |
+
)
|
205 |
+
with gr.Row():
|
206 |
+
with gr.Column():
|
207 |
+
prior_guidance_scale = gr.Slider(
|
208 |
+
label="Prior Guidance Scale (CFG)",
|
209 |
+
minimum=0,
|
210 |
+
maximum=20,
|
211 |
+
step=0.1,
|
212 |
+
value=4.0,
|
213 |
+
)
|
214 |
+
with gr.Column():
|
215 |
+
decoder_guidance_scale = gr.Slider(
|
216 |
+
label="Decoder Guidance Scale (CFG)",
|
217 |
+
minimum=0,
|
218 |
+
maximum=20,
|
219 |
+
step=0.1,
|
220 |
+
value=0.0,
|
221 |
+
)
|
222 |
+
with gr.Row():
|
223 |
+
with gr.Column():
|
224 |
+
prior_num_inference_steps = gr.Slider(
|
225 |
+
label="Prior Inference Steps",
|
226 |
+
minimum=1,
|
227 |
+
maximum=100,
|
228 |
+
step=1,
|
229 |
+
value=20,
|
230 |
+
)
|
231 |
+
with gr.Column():
|
232 |
+
decoder_num_inference_steps = gr.Slider(
|
233 |
+
label="Decoder Inference Steps",
|
234 |
+
minimum=1,
|
235 |
+
maximum=100,
|
236 |
+
step=1,
|
237 |
+
value=20,
|
238 |
+
)
|
239 |
+
|
240 |
+
with gr.Column():
|
241 |
+
result = gr.Gallery(label="Result", show_label=False, height=768)
|
242 |
|
243 |
inputs = [
|
244 |
prompt,
|
|
|
252 |
decoder_num_inference_steps,
|
253 |
# decoder_timesteps,
|
254 |
decoder_guidance_scale,
|
255 |
+
batch_size_per_prompt,
|
256 |
+
number_of_images_per_prompt
|
257 |
]
|
258 |
gr.on(
|
259 |
triggers=[prompt.submit, negative_prompt.submit, run_button.click],
|
|
|
268 |
outputs=result,
|
269 |
api_name="run",
|
270 |
)
|
271 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
272 |
if __name__ == "__main__":
|
273 |
+
app.queue().launch(share=share,inbrowser=True)
|