Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -1,14 +1,44 @@
|
|
1 |
import json
|
2 |
import random
|
3 |
import requests
|
4 |
-
import os
|
5 |
-
from PIL import Image
|
6 |
-
|
7 |
import gradio as gr
|
8 |
import numpy as np
|
9 |
import spaces
|
10 |
import torch
|
11 |
from diffusers import DiffusionPipeline, LCMScheduler
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
def get_image(image_data):
|
14 |
if isinstance(image_data, str):
|
@@ -21,20 +51,22 @@ def get_image(image_data):
|
|
21 |
print(f"Unexpected image_data format: {type(image_data)}")
|
22 |
return None
|
23 |
|
|
|
24 |
if local_path and os.path.exists(local_path):
|
25 |
try:
|
26 |
-
Image.open(local_path).verify()
|
27 |
return local_path
|
28 |
except Exception as e:
|
29 |
print(f"Error loading local image {local_path}: {e}")
|
30 |
|
|
|
31 |
if hf_url:
|
32 |
try:
|
33 |
response = requests.get(hf_url)
|
34 |
if response.status_code == 200:
|
35 |
img = Image.open(requests.get(hf_url, stream=True).raw)
|
36 |
-
img.verify()
|
37 |
-
img.save(local_path)
|
38 |
return local_path
|
39 |
else:
|
40 |
print(f"Failed to fetch image from URL {hf_url}. Status code: {response.status_code}")
|
@@ -44,51 +76,6 @@ def get_image(image_data):
|
|
44 |
print(f"Failed to load image for {image_data}")
|
45 |
return None
|
46 |
|
47 |
-
with open("sdxl_lora.json", "r") as file:
|
48 |
-
data = json.load(file)
|
49 |
-
sdxl_loras_raw = [
|
50 |
-
{
|
51 |
-
"image": get_image(item["image"]),
|
52 |
-
"title": item["title"],
|
53 |
-
"repo": item["repo"],
|
54 |
-
"trigger_word": item["trigger_word"],
|
55 |
-
"weights": item["weights"],
|
56 |
-
"is_pivotal": item.get("is_pivotal", False),
|
57 |
-
"text_embedding_weights": item.get("text_embedding_weights", None),
|
58 |
-
"likes": item.get("likes", 0),
|
59 |
-
}
|
60 |
-
for item in data
|
61 |
-
]
|
62 |
-
|
63 |
-
sdxl_loras_raw = sorted(sdxl_loras_raw, key=lambda x: x["likes"], reverse=True)
|
64 |
-
|
65 |
-
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
66 |
-
model_id = "stabilityai/stable-diffusion-xl-base-1.0"
|
67 |
-
|
68 |
-
pipe = DiffusionPipeline.from_pretrained(model_id, variant="fp16")
|
69 |
-
|
70 |
-
# Create LCMScheduler with default config
|
71 |
-
lcm_scheduler = LCMScheduler.from_config(pipe.scheduler.config)
|
72 |
-
|
73 |
-
# Remove the 'skip_prk_steps' if it exists in the config
|
74 |
-
if hasattr(lcm_scheduler.config, 'skip_prk_steps'):
|
75 |
-
delattr(lcm_scheduler.config, 'skip_prk_steps')
|
76 |
-
|
77 |
-
pipe.scheduler = lcm_scheduler
|
78 |
-
pipe.to(device=DEVICE, dtype=torch.float16)
|
79 |
-
|
80 |
-
# Load Flash SDXL LoRA
|
81 |
-
flash_sdxl_id = "jasperai/flash-sdxl"
|
82 |
-
pipe.load_lora_weights(flash_sdxl_id, adapter_name="flash_lora")
|
83 |
-
|
84 |
-
MAX_SEED = np.iinfo(np.int32).max
|
85 |
-
MAX_IMAGE_SIZE = 1024
|
86 |
-
|
87 |
-
def update_selection(selected_state: gr.SelectData, gr_sdxl_loras):
|
88 |
-
lora_id = gr_sdxl_loras[selected_state.index]["repo"]
|
89 |
-
trigger_word = gr_sdxl_loras[selected_state.index]["trigger_word"]
|
90 |
-
return lora_id, trigger_word
|
91 |
-
|
92 |
@spaces.GPU
|
93 |
def infer(
|
94 |
pre_prompt,
|
@@ -102,168 +89,150 @@ def infer(
|
|
102 |
user_lora_weight,
|
103 |
progress=gr.Progress(track_tqdm=True),
|
104 |
):
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
prompt=prompt,
|
125 |
-
negative_prompt=negative_prompt,
|
126 |
-
guidance_scale=1.0, # Flash Diffusion typically uses guidance_scale=1
|
127 |
-
num_inference_steps=4, # Flash Diffusion uses fewer steps
|
128 |
-
generator=generator,
|
129 |
-
).images[0]
|
130 |
-
|
131 |
-
return image
|
132 |
-
except Exception as e:
|
133 |
-
gr.Error(f"An error occurred: {str(e)}")
|
134 |
-
return None
|
135 |
|
136 |
css = """
|
137 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
138 |
text-align: center;
|
139 |
-
display:block;
|
140 |
}
|
141 |
-
|
142 |
-
|
143 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
}
|
145 |
"""
|
146 |
|
147 |
with gr.Blocks(css=css) as demo:
|
148 |
gr.Markdown(
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
|
|
|
|
159 |
)
|
160 |
|
161 |
gr_sdxl_loras = gr.State(value=sdxl_loras_raw)
|
162 |
gr_lora_id = gr.State(value="")
|
163 |
|
164 |
with gr.Row():
|
165 |
-
with gr.
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
172 |
|
173 |
-
user_lora_weight = gr.Slider(
|
174 |
-
label="Selected LoRA Weight",
|
175 |
-
minimum=0.5,
|
176 |
-
maximum=3,
|
177 |
-
step=0.1,
|
178 |
-
value=1,
|
179 |
-
)
|
180 |
-
|
181 |
-
gallery = gr.Gallery(
|
182 |
-
value=[(item["image"], item["title"]) for item in sdxl_loras_raw],
|
183 |
-
label="SDXL LoRA Gallery",
|
184 |
-
allow_preview=False,
|
185 |
-
columns=3,
|
186 |
-
elem_id="gallery",
|
187 |
-
show_share_button=False,
|
188 |
-
)
|
189 |
-
|
190 |
-
with gr.Column():
|
191 |
with gr.Row():
|
192 |
-
|
193 |
-
|
194 |
-
show_label=False,
|
195 |
-
max_lines=1,
|
196 |
-
placeholder="Enter your prompt",
|
197 |
-
container=False,
|
198 |
-
scale=5,
|
199 |
-
)
|
200 |
|
201 |
-
|
202 |
-
|
203 |
-
result = gr.Image(label="Result", show_label=False)
|
204 |
|
205 |
with gr.Accordion("Advanced Settings", open=False):
|
206 |
-
pre_prompt = gr.
|
207 |
label="Pre-Prompt",
|
208 |
-
show_label=True,
|
209 |
-
max_lines=1,
|
210 |
placeholder="Pre Prompt from the LoRA config",
|
211 |
-
|
212 |
-
scale=5,
|
213 |
)
|
214 |
|
215 |
-
seed = gr.Slider(
|
216 |
-
label="Seed",
|
217 |
-
minimum=0,
|
218 |
-
maximum=MAX_SEED,
|
219 |
-
step=1,
|
220 |
-
value=0,
|
221 |
-
)
|
222 |
-
|
223 |
-
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
224 |
-
|
225 |
with gr.Row():
|
226 |
-
|
227 |
-
label="
|
228 |
-
minimum=
|
229 |
-
maximum=
|
230 |
step=1,
|
231 |
-
value=
|
232 |
)
|
|
|
233 |
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
)
|
242 |
|
243 |
-
|
244 |
-
"
|
245 |
-
|
246 |
-
|
|
|
|
|
247 |
)
|
248 |
|
249 |
-
negative_prompt = gr.
|
250 |
label="Negative Prompt",
|
251 |
-
show_label=False,
|
252 |
-
max_lines=1,
|
253 |
placeholder="Enter a negative Prompt",
|
254 |
-
|
255 |
)
|
256 |
|
257 |
gr.on(
|
258 |
-
[
|
259 |
-
run_button.click,
|
260 |
-
seed.change,
|
261 |
-
randomize_seed.change,
|
262 |
-
prompt.submit,
|
263 |
-
negative_prompt.change,
|
264 |
-
negative_prompt.submit,
|
265 |
-
guidance_scale.change,
|
266 |
-
],
|
267 |
fn=infer,
|
268 |
inputs=[
|
269 |
pre_prompt,
|
@@ -274,24 +243,30 @@ with gr.Blocks(css=css) as demo:
|
|
274 |
negative_prompt,
|
275 |
guidance_scale,
|
276 |
user_lora_selector,
|
277 |
-
|
278 |
],
|
279 |
outputs=[result],
|
280 |
)
|
281 |
|
|
|
|
|
282 |
gallery.select(
|
283 |
fn=update_selection,
|
284 |
inputs=[gr_sdxl_loras],
|
285 |
-
outputs=[
|
286 |
-
user_lora_selector,
|
287 |
-
pre_prompt,
|
288 |
-
],
|
289 |
-
show_progress="hidden",
|
290 |
)
|
291 |
|
292 |
-
gr.Markdown("**Disclaimer:**")
|
293 |
gr.Markdown(
|
294 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
295 |
)
|
296 |
|
297 |
demo.queue().launch()
|
|
|
1 |
import json
|
2 |
import random
|
3 |
import requests
|
|
|
|
|
|
|
4 |
import gradio as gr
|
5 |
import numpy as np
|
6 |
import spaces
|
7 |
import torch
|
8 |
from diffusers import DiffusionPipeline, LCMScheduler
|
9 |
+
from PIL import Image
|
10 |
+
import os
|
11 |
+
|
12 |
+
# Load the JSON data
|
13 |
+
with open("sdxl_lora.json", "r") as file:
|
14 |
+
data = json.load(file)
|
15 |
+
sdxl_loras_raw = sorted(data, key=lambda x: x["likes"], reverse=True)
|
16 |
+
|
17 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
18 |
+
model_id = "stabilityai/stable-diffusion-xl-base-1.0"
|
19 |
+
|
20 |
+
pipe = DiffusionPipeline.from_pretrained(model_id, variant="fp16")
|
21 |
+
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
|
22 |
+
pipe.to(device=DEVICE, dtype=torch.float16)
|
23 |
+
|
24 |
+
MAX_SEED = np.iinfo(np.int32).max
|
25 |
+
MAX_IMAGE_SIZE = 1024
|
26 |
+
|
27 |
+
def update_selection(selected_state: gr.SelectData, gr_sdxl_loras):
|
28 |
+
lora_id = gr_sdxl_loras[selected_state.index]["repo"]
|
29 |
+
trigger_word = gr_sdxl_loras[selected_state.index]["trigger_word"]
|
30 |
+
return lora_id, trigger_word
|
31 |
+
|
32 |
+
def load_lora_for_style(style_repo):
|
33 |
+
pipe.unload_lora_weights()
|
34 |
+
pipe.load_lora_weights("jasperai/flash-sdxl", adapter_name="lora")
|
35 |
+
pipe.set_adapters(["lora", new_adapter_id], adapter_weights=[1.0, user_lora_weight])
|
36 |
+
|
37 |
+
if new_adapter_id not in loaded_adapters["unet"]:
|
38 |
+
gr.Info("Swapping LoRA")
|
39 |
+
pipe.unload_lora_weights()
|
40 |
+
pipe.load_lora_weights(flash_sdxl_id, adapter_name="lora")
|
41 |
+
pipe.load_lora_weights(user_lora_selector, adapter_name=new_adapter_id)
|
42 |
|
43 |
def get_image(image_data):
|
44 |
if isinstance(image_data, str):
|
|
|
51 |
print(f"Unexpected image_data format: {type(image_data)}")
|
52 |
return None
|
53 |
|
54 |
+
# Try loading from local path first
|
55 |
if local_path and os.path.exists(local_path):
|
56 |
try:
|
57 |
+
Image.open(local_path).verify() # Verify that it's a valid image
|
58 |
return local_path
|
59 |
except Exception as e:
|
60 |
print(f"Error loading local image {local_path}: {e}")
|
61 |
|
62 |
+
# If local path fails or doesn't exist, try URL
|
63 |
if hf_url:
|
64 |
try:
|
65 |
response = requests.get(hf_url)
|
66 |
if response.status_code == 200:
|
67 |
img = Image.open(requests.get(hf_url, stream=True).raw)
|
68 |
+
img.verify() # Verify that it's a valid image
|
69 |
+
img.save(local_path) # Save for future use
|
70 |
return local_path
|
71 |
else:
|
72 |
print(f"Failed to fetch image from URL {hf_url}. Status code: {response.status_code}")
|
|
|
76 |
print(f"Failed to load image for {image_data}")
|
77 |
return None
|
78 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
@spaces.GPU
|
80 |
def infer(
|
81 |
pre_prompt,
|
|
|
89 |
user_lora_weight,
|
90 |
progress=gr.Progress(track_tqdm=True),
|
91 |
):
|
92 |
+
load_lora_for_style(user_lora_selector)
|
93 |
+
|
94 |
+
if randomize_seed:
|
95 |
+
seed = random.randint(0, MAX_SEED)
|
96 |
+
|
97 |
+
generator = torch.Generator().manual_seed(seed)
|
98 |
+
|
99 |
+
if pre_prompt != "":
|
100 |
+
prompt = f"{pre_prompt} {prompt}"
|
101 |
+
|
102 |
+
image = pipe(
|
103 |
+
prompt=prompt,
|
104 |
+
negative_prompt=negative_prompt,
|
105 |
+
guidance_scale=guidance_scale,
|
106 |
+
num_inference_steps=num_inference_steps,
|
107 |
+
generator=generator,
|
108 |
+
).images[0]
|
109 |
+
|
110 |
+
return image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
111 |
|
112 |
css = """
|
113 |
+
body {
|
114 |
+
background-color: #1a1a1a;
|
115 |
+
color: #ffffff;
|
116 |
+
}
|
117 |
+
.container {
|
118 |
+
max-width: 900px;
|
119 |
+
margin: auto;
|
120 |
+
padding: 20px;
|
121 |
+
}
|
122 |
+
h1, h2 {
|
123 |
+
color: #4CAF50;
|
124 |
text-align: center;
|
|
|
125 |
}
|
126 |
+
.gallery {
|
127 |
+
display: flex;
|
128 |
+
flex-wrap: wrap;
|
129 |
+
justify-content: center;
|
130 |
+
}
|
131 |
+
.gallery img {
|
132 |
+
margin: 10px;
|
133 |
+
border-radius: 10px;
|
134 |
+
transition: transform 0.3s ease;
|
135 |
+
}
|
136 |
+
.gallery img:hover {
|
137 |
+
transform: scale(1.05);
|
138 |
+
}
|
139 |
+
.gradio-slider input[type="range"] {
|
140 |
+
background-color: #4CAF50;
|
141 |
+
}
|
142 |
+
.gradio-button {
|
143 |
+
background-color: #4CAF50 !important;
|
144 |
}
|
145 |
"""
|
146 |
|
147 |
with gr.Blocks(css=css) as demo:
|
148 |
gr.Markdown(
|
149 |
+
"""
|
150 |
+
# β‘ FlashDiffusion: Araminta K's FlashLoRA Showcase β‘
|
151 |
+
|
152 |
+
This interactive demo showcases [Araminta K's models](https://huggingface.co/alvdansen) using [Flash Diffusion](https://gojasper.github.io/flash-diffusion-project/) technology.
|
153 |
+
|
154 |
+
## Acknowledgments
|
155 |
+
- Original Flash Diffusion technology by the Jasper AI team
|
156 |
+
- Based on the paper: [Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation](http://arxiv.org/abs/2406.02347) by ClΓ©ment Chadebec, Onur Tasar, Eyal Benaroche and Benjamin Aubin
|
157 |
+
- Models showcased here are created by Araminta K at Alvdansen Labs
|
158 |
+
|
159 |
+
Explore the power of FlashLoRA with Araminta K's unique artistic styles!
|
160 |
+
"""
|
161 |
)
|
162 |
|
163 |
gr_sdxl_loras = gr.State(value=sdxl_loras_raw)
|
164 |
gr_lora_id = gr.State(value="")
|
165 |
|
166 |
with gr.Row():
|
167 |
+
with gr.Column(scale=2):
|
168 |
+
gallery = gr.Gallery(
|
169 |
+
value=[(get_image(item["image"]), item["title"]) for item in sdxl_loras_raw if get_image(item["image"]) is not None],
|
170 |
+
label="SDXL LoRA Gallery",
|
171 |
+
show_label=False,
|
172 |
+
elem_id="gallery",
|
173 |
+
columns=3,
|
174 |
+
height=600,
|
175 |
+
)
|
176 |
+
|
177 |
+
user_lora_selector = gr.Textbox(
|
178 |
+
label="Current Selected LoRA",
|
179 |
+
interactive=False,
|
180 |
+
)
|
181 |
+
|
182 |
+
with gr.Column(scale=3):
|
183 |
+
prompt = gr.Textbox(
|
184 |
+
label="Prompt",
|
185 |
+
placeholder="Enter your prompt",
|
186 |
+
lines=3,
|
187 |
+
)
|
188 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
189 |
with gr.Row():
|
190 |
+
run_button = gr.Button("Run", variant="primary")
|
191 |
+
clear_button = gr.Button("Clear")
|
|
|
|
|
|
|
|
|
|
|
|
|
192 |
|
193 |
+
result = gr.Image(label="Result", height=512)
|
|
|
|
|
194 |
|
195 |
with gr.Accordion("Advanced Settings", open=False):
|
196 |
+
pre_prompt = gr.Textbox(
|
197 |
label="Pre-Prompt",
|
|
|
|
|
198 |
placeholder="Pre Prompt from the LoRA config",
|
199 |
+
lines=2,
|
|
|
200 |
)
|
201 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
202 |
with gr.Row():
|
203 |
+
seed = gr.Slider(
|
204 |
+
label="Seed",
|
205 |
+
minimum=0,
|
206 |
+
maximum=MAX_SEED,
|
207 |
step=1,
|
208 |
+
value=0,
|
209 |
)
|
210 |
+
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
211 |
|
212 |
+
num_inference_steps = gr.Slider(
|
213 |
+
label="Number of inference steps",
|
214 |
+
minimum=4,
|
215 |
+
maximum=8,
|
216 |
+
step=1,
|
217 |
+
value=4,
|
218 |
+
)
|
|
|
219 |
|
220 |
+
guidance_scale = gr.Slider(
|
221 |
+
label="Guidance Scale",
|
222 |
+
minimum=1,
|
223 |
+
maximum=6,
|
224 |
+
step=0.5,
|
225 |
+
value=1,
|
226 |
)
|
227 |
|
228 |
+
negative_prompt = gr.Textbox(
|
229 |
label="Negative Prompt",
|
|
|
|
|
230 |
placeholder="Enter a negative Prompt",
|
231 |
+
lines=2,
|
232 |
)
|
233 |
|
234 |
gr.on(
|
235 |
+
[run_button.click, prompt.submit],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
236 |
fn=infer,
|
237 |
inputs=[
|
238 |
pre_prompt,
|
|
|
243 |
negative_prompt,
|
244 |
guidance_scale,
|
245 |
user_lora_selector,
|
246 |
+
gr.Slider(label="Selected LoRA Weight", minimum=0.5, maximum=3, step=0.1, value=1),
|
247 |
],
|
248 |
outputs=[result],
|
249 |
)
|
250 |
|
251 |
+
clear_button.click(lambda: "", outputs=[prompt, result])
|
252 |
+
|
253 |
gallery.select(
|
254 |
fn=update_selection,
|
255 |
inputs=[gr_sdxl_loras],
|
256 |
+
outputs=[user_lora_selector, pre_prompt],
|
|
|
|
|
|
|
|
|
257 |
)
|
258 |
|
|
|
259 |
gr.Markdown(
|
260 |
+
"""
|
261 |
+
## Unleash Your Creativity!
|
262 |
+
|
263 |
+
This showcase brings together the speed of Flash Diffusion and the artistic flair of Araminta K's models.
|
264 |
+
Craft your prompts, adjust the settings, and watch as AI brings your ideas to life in stunning detail.
|
265 |
+
|
266 |
+
Remember to use this tool ethically and respect copyright and individual privacy.
|
267 |
+
|
268 |
+
Enjoy exploring these unique artistic styles!
|
269 |
+
"""
|
270 |
)
|
271 |
|
272 |
demo.queue().launch()
|