Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -4,44 +4,99 @@ import numpy as np
|
|
4 |
import PIL.Image
|
5 |
from PIL import Image
|
6 |
import random
|
7 |
-
from diffusers import
|
8 |
-
from diffusers import
|
9 |
-
import cv2
|
10 |
import torch
|
|
|
11 |
|
12 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
13 |
|
|
|
14 |
pipe = StableDiffusionXLPipeline.from_pretrained(
|
15 |
"votepurchase/noobreal_v21",
|
16 |
torch_dtype=torch.float16,
|
|
|
|
|
17 |
)
|
18 |
|
19 |
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
|
20 |
pipe.to(device)
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
MAX_SEED = np.iinfo(np.int32).max
|
23 |
MAX_IMAGE_SIZE = 1216
|
24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
@spaces.GPU
|
27 |
def infer(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps):
|
28 |
-
|
|
|
|
|
29 |
if randomize_seed:
|
30 |
seed = random.randint(0, MAX_SEED)
|
31 |
|
32 |
-
generator = torch.Generator().manual_seed(seed)
|
33 |
-
|
34 |
-
|
35 |
-
prompt
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
|
46 |
|
47 |
css = """
|
@@ -60,7 +115,7 @@ with gr.Blocks(css=css) as demo:
|
|
60 |
label="Prompt",
|
61 |
show_label=False,
|
62 |
max_lines=1,
|
63 |
-
placeholder="Enter your prompt",
|
64 |
container=False,
|
65 |
)
|
66 |
|
@@ -93,7 +148,7 @@ with gr.Blocks(css=css) as demo:
|
|
93 |
minimum=256,
|
94 |
maximum=MAX_IMAGE_SIZE,
|
95 |
step=32,
|
96 |
-
value=1024
|
97 |
)
|
98 |
|
99 |
height = gr.Slider(
|
@@ -101,7 +156,7 @@ with gr.Blocks(css=css) as demo:
|
|
101 |
minimum=256,
|
102 |
maximum=MAX_IMAGE_SIZE,
|
103 |
step=32,
|
104 |
-
value=1024
|
105 |
)
|
106 |
|
107 |
with gr.Row():
|
@@ -121,7 +176,7 @@ with gr.Blocks(css=css) as demo:
|
|
121 |
value=28,
|
122 |
)
|
123 |
|
124 |
-
run_button.click(
|
125 |
fn=infer,
|
126 |
inputs=[prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
|
127 |
outputs=[result]
|
|
|
4 |
import PIL.Image
|
5 |
from PIL import Image
|
6 |
import random
|
7 |
+
from diffusers import StableDiffusionXLPipeline
|
8 |
+
from diffusers import EulerAncestralDiscreteScheduler
|
|
|
9 |
import torch
|
10 |
+
from compel import Compel, ReturnedEmbeddingsType
|
11 |
|
12 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
13 |
|
14 |
+
# Make sure to use torch.float16 consistently throughout the pipeline
|
15 |
pipe = StableDiffusionXLPipeline.from_pretrained(
|
16 |
"votepurchase/noobreal_v21",
|
17 |
torch_dtype=torch.float16,
|
18 |
+
variant="fp16", # Explicitly use fp16 variant
|
19 |
+
use_safetensors=True # Use safetensors if available
|
20 |
)
|
21 |
|
22 |
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
|
23 |
pipe.to(device)
|
24 |
|
25 |
+
# Force all components to use the same dtype
|
26 |
+
pipe.text_encoder.to(torch.float16)
|
27 |
+
pipe.text_encoder_2.to(torch.float16)
|
28 |
+
pipe.vae.to(torch.float16)
|
29 |
+
pipe.unet.to(torch.float16)
|
30 |
+
|
31 |
+
# 追加: Initialize Compel for long prompt processing
|
32 |
+
compel = Compel(
|
33 |
+
tokenizer=[pipe.tokenizer, pipe.tokenizer_2],
|
34 |
+
text_encoder=[pipe.text_encoder, pipe.text_encoder_2],
|
35 |
+
returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
|
36 |
+
requires_pooled=[False, True],
|
37 |
+
truncate_long_prompts=False
|
38 |
+
)
|
39 |
+
|
40 |
MAX_SEED = np.iinfo(np.int32).max
|
41 |
MAX_IMAGE_SIZE = 1216
|
42 |
|
43 |
+
# 追加: Simple long prompt processing function
|
44 |
+
def process_long_prompt(prompt, negative_prompt=""):
|
45 |
+
"""Simple long prompt processing using Compel"""
|
46 |
+
try:
|
47 |
+
conditioning, pooled = compel([prompt, negative_prompt])
|
48 |
+
return conditioning, pooled
|
49 |
+
except Exception as e:
|
50 |
+
print(f"Long prompt processing failed: {e}, falling back to standard processing")
|
51 |
+
return None, None
|
52 |
|
53 |
@spaces.GPU
|
54 |
def infer(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps):
|
55 |
+
# 変更: Remove the 60-word limit warning and add long prompt check
|
56 |
+
use_long_prompt = len(prompt.split()) > 60 or len(prompt) > 300
|
57 |
+
|
58 |
if randomize_seed:
|
59 |
seed = random.randint(0, MAX_SEED)
|
60 |
|
61 |
+
generator = torch.Generator(device=device).manual_seed(seed)
|
62 |
+
|
63 |
+
try:
|
64 |
+
# 追加: Try long prompt processing first if prompt is long
|
65 |
+
if use_long_prompt:
|
66 |
+
print("Using long prompt processing...")
|
67 |
+
conditioning, pooled = process_long_prompt(prompt, negative_prompt)
|
68 |
+
|
69 |
+
if conditioning is not None:
|
70 |
+
output_image = pipe(
|
71 |
+
prompt_embeds=conditioning[0:1],
|
72 |
+
pooled_prompt_embeds=pooled[0:1],
|
73 |
+
negative_prompt_embeds=conditioning[1:2],
|
74 |
+
negative_pooled_prompt_embeds=pooled[1:2],
|
75 |
+
guidance_scale=guidance_scale,
|
76 |
+
num_inference_steps=num_inference_steps,
|
77 |
+
width=width,
|
78 |
+
height=height,
|
79 |
+
generator=generator
|
80 |
+
).images[0]
|
81 |
+
return output_image
|
82 |
+
|
83 |
+
# Fall back to standard processing
|
84 |
+
output_image = pipe(
|
85 |
+
prompt=prompt,
|
86 |
+
negative_prompt=negative_prompt,
|
87 |
+
guidance_scale=guidance_scale,
|
88 |
+
num_inference_steps=num_inference_steps,
|
89 |
+
width=width,
|
90 |
+
height=height,
|
91 |
+
generator=generator
|
92 |
+
).images[0]
|
93 |
+
|
94 |
+
return output_image
|
95 |
+
except RuntimeError as e:
|
96 |
+
print(f"Error during generation: {e}")
|
97 |
+
# Return a blank image with error message
|
98 |
+
error_img = Image.new('RGB', (width, height), color=(0, 0, 0))
|
99 |
+
return error_img
|
100 |
|
101 |
|
102 |
css = """
|
|
|
115 |
label="Prompt",
|
116 |
show_label=False,
|
117 |
max_lines=1,
|
118 |
+
placeholder="Enter your prompt (long prompts are automatically supported)", # 変更: Updated placeholder
|
119 |
container=False,
|
120 |
)
|
121 |
|
|
|
148 |
minimum=256,
|
149 |
maximum=MAX_IMAGE_SIZE,
|
150 |
step=32,
|
151 |
+
value=1024,
|
152 |
)
|
153 |
|
154 |
height = gr.Slider(
|
|
|
156 |
minimum=256,
|
157 |
maximum=MAX_IMAGE_SIZE,
|
158 |
step=32,
|
159 |
+
value=1024,
|
160 |
)
|
161 |
|
162 |
with gr.Row():
|
|
|
176 |
value=28,
|
177 |
)
|
178 |
|
179 |
+
run_button.click(
|
180 |
fn=infer,
|
181 |
inputs=[prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
|
182 |
outputs=[result]
|