Spaces:
Running
on
Zero
Running
on
Zero
Update
Browse files
app.py
CHANGED
@@ -7,16 +7,15 @@ import random
|
|
7 |
|
8 |
import gradio as gr
|
9 |
import numpy as np
|
|
|
10 |
import torch
|
11 |
-
|
12 |
-
from model import Model
|
13 |
|
14 |
DESCRIPTION = "# [UniDiffuser](https://github.com/thu-ml/unidiffuser)"
|
15 |
|
16 |
if not torch.cuda.is_available():
|
17 |
DESCRIPTION += "\n<p>Running on CPU 🥶</p>"
|
18 |
|
19 |
-
model = Model()
|
20 |
|
21 |
MAX_SEED = np.iinfo(np.int32).max
|
22 |
|
@@ -27,6 +26,67 @@ def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
|
|
27 |
return seed
|
28 |
|
29 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
def create_demo(mode_name: str) -> gr.Blocks:
|
31 |
with gr.Blocks() as demo:
|
32 |
with gr.Row():
|
@@ -82,7 +142,7 @@ def create_demo(mode_name: str) -> gr.Blocks:
|
|
82 |
outputs=seed,
|
83 |
queue=False,
|
84 |
).then(
|
85 |
-
fn=
|
86 |
inputs=[
|
87 |
mode,
|
88 |
prompt,
|
|
|
7 |
|
8 |
import gradio as gr
|
9 |
import numpy as np
|
10 |
+
import PIL.Image
|
11 |
import torch
|
12 |
+
from diffusers import UniDiffuserPipeline
|
|
|
13 |
|
14 |
DESCRIPTION = "# [UniDiffuser](https://github.com/thu-ml/unidiffuser)"
|
15 |
|
16 |
if not torch.cuda.is_available():
|
17 |
DESCRIPTION += "\n<p>Running on CPU 🥶</p>"
|
18 |
|
|
|
19 |
|
20 |
MAX_SEED = np.iinfo(np.int32).max
|
21 |
|
|
|
26 |
return seed
|
27 |
|
28 |
|
29 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
30 |
+
if torch.cuda.is_available():
|
31 |
+
pipe = UniDiffuserPipeline.from_pretrained("thu-ml/unidiffuser-v1", torch_dtype=torch.float16)
|
32 |
+
pipe.to(device)
|
33 |
+
|
34 |
+
|
35 |
+
def run(
|
36 |
+
mode: str,
|
37 |
+
prompt: str,
|
38 |
+
image: PIL.Image.Image | None,
|
39 |
+
seed: int = 0,
|
40 |
+
num_steps: int = 20,
|
41 |
+
guidance_scale: float = 8.0,
|
42 |
+
) -> tuple[PIL.Image.Image | None, str]:
|
43 |
+
generator = torch.Generator(device=device).manual_seed(seed)
|
44 |
+
if mode == "t2i":
|
45 |
+
pipe.set_text_to_image_mode()
|
46 |
+
sample = pipe(prompt=prompt, num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
47 |
+
return sample.images[0], ""
|
48 |
+
elif mode == "i2t":
|
49 |
+
pipe.set_image_to_text_mode()
|
50 |
+
sample = pipe(image=image, num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
51 |
+
return None, sample.text[0]
|
52 |
+
elif mode == "joint":
|
53 |
+
pipe.set_joint_mode()
|
54 |
+
sample = pipe(num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
55 |
+
return sample.images[0], sample.text[0]
|
56 |
+
elif mode == "i":
|
57 |
+
pipe.set_image_mode()
|
58 |
+
sample = pipe(num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
59 |
+
return sample.images[0], ""
|
60 |
+
elif mode == "t":
|
61 |
+
pipe.set_text_mode()
|
62 |
+
sample = pipe(num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
63 |
+
return None, sample.text[0]
|
64 |
+
elif mode == "i2t2i":
|
65 |
+
pipe.set_image_to_text_mode()
|
66 |
+
sample = pipe(image=image, num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
67 |
+
pipe.set_text_to_image_mode()
|
68 |
+
sample = pipe(
|
69 |
+
prompt=sample.text[0],
|
70 |
+
num_inference_steps=num_steps,
|
71 |
+
guidance_scale=guidance_scale,
|
72 |
+
generator=generator,
|
73 |
+
)
|
74 |
+
return sample.images[0], ""
|
75 |
+
elif mode == "t2i2t":
|
76 |
+
pipe.set_text_to_image_mode()
|
77 |
+
sample = pipe(prompt=prompt, num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
78 |
+
pipe.set_image_to_text_mode()
|
79 |
+
sample = pipe(
|
80 |
+
image=sample.images[0],
|
81 |
+
num_inference_steps=num_steps,
|
82 |
+
guidance_scale=guidance_scale,
|
83 |
+
generator=generator,
|
84 |
+
)
|
85 |
+
return None, sample.text[0]
|
86 |
+
else:
|
87 |
+
raise ValueError
|
88 |
+
|
89 |
+
|
90 |
def create_demo(mode_name: str) -> gr.Blocks:
|
91 |
with gr.Blocks() as demo:
|
92 |
with gr.Row():
|
|
|
142 |
outputs=seed,
|
143 |
queue=False,
|
144 |
).then(
|
145 |
+
fn=run,
|
146 |
inputs=[
|
147 |
mode,
|
148 |
prompt,
|
model.py
DELETED
@@ -1,78 +0,0 @@
|
|
1 |
-
from __future__ import annotations
|
2 |
-
|
3 |
-
import PIL.Image
|
4 |
-
import torch
|
5 |
-
from diffusers import UniDiffuserPipeline
|
6 |
-
|
7 |
-
|
8 |
-
class Model:
|
9 |
-
def __init__(self):
|
10 |
-
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
11 |
-
if self.device.type == "cuda":
|
12 |
-
self.pipe = UniDiffuserPipeline.from_pretrained("thu-ml/unidiffuser-v1", torch_dtype=torch.float16)
|
13 |
-
self.pipe.to(self.device)
|
14 |
-
else:
|
15 |
-
self.pipe = UniDiffuserPipeline.from_pretrained("thu-ml/unidiffuser-v1")
|
16 |
-
|
17 |
-
def run(
|
18 |
-
self,
|
19 |
-
mode: str,
|
20 |
-
prompt: str,
|
21 |
-
image: PIL.Image.Image | None,
|
22 |
-
seed: int = 0,
|
23 |
-
num_steps: int = 20,
|
24 |
-
guidance_scale: float = 8.0,
|
25 |
-
) -> tuple[PIL.Image.Image | None, str]:
|
26 |
-
generator = torch.Generator(device=self.device).manual_seed(seed)
|
27 |
-
if mode == "t2i":
|
28 |
-
self.pipe.set_text_to_image_mode()
|
29 |
-
sample = self.pipe(
|
30 |
-
prompt=prompt, num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator
|
31 |
-
)
|
32 |
-
return sample.images[0], ""
|
33 |
-
elif mode == "i2t":
|
34 |
-
self.pipe.set_image_to_text_mode()
|
35 |
-
sample = self.pipe(
|
36 |
-
image=image, num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator
|
37 |
-
)
|
38 |
-
return None, sample.text[0]
|
39 |
-
elif mode == "joint":
|
40 |
-
self.pipe.set_joint_mode()
|
41 |
-
sample = self.pipe(num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
42 |
-
return sample.images[0], sample.text[0]
|
43 |
-
elif mode == "i":
|
44 |
-
self.pipe.set_image_mode()
|
45 |
-
sample = self.pipe(num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
46 |
-
return sample.images[0], ""
|
47 |
-
elif mode == "t":
|
48 |
-
self.pipe.set_text_mode()
|
49 |
-
sample = self.pipe(num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator)
|
50 |
-
return None, sample.text[0]
|
51 |
-
elif mode == "i2t2i":
|
52 |
-
self.pipe.set_image_to_text_mode()
|
53 |
-
sample = self.pipe(
|
54 |
-
image=image, num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator
|
55 |
-
)
|
56 |
-
self.pipe.set_text_to_image_mode()
|
57 |
-
sample = self.pipe(
|
58 |
-
prompt=sample.text[0],
|
59 |
-
num_inference_steps=num_steps,
|
60 |
-
guidance_scale=guidance_scale,
|
61 |
-
generator=generator,
|
62 |
-
)
|
63 |
-
return sample.images[0], ""
|
64 |
-
elif mode == "t2i2t":
|
65 |
-
self.pipe.set_text_to_image_mode()
|
66 |
-
sample = self.pipe(
|
67 |
-
prompt=prompt, num_inference_steps=num_steps, guidance_scale=guidance_scale, generator=generator
|
68 |
-
)
|
69 |
-
self.pipe.set_image_to_text_mode()
|
70 |
-
sample = self.pipe(
|
71 |
-
image=sample.images[0],
|
72 |
-
num_inference_steps=num_steps,
|
73 |
-
guidance_scale=guidance_scale,
|
74 |
-
generator=generator,
|
75 |
-
)
|
76 |
-
return None, sample.text[0]
|
77 |
-
else:
|
78 |
-
raise ValueError
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|