Spaces:
Runtime error
Runtime error
Update
Browse files
app.py
CHANGED
@@ -31,6 +31,7 @@ def run(
|
|
31 |
image: PIL.Image.Image,
|
32 |
prompt: str,
|
33 |
negative_prompt: str,
|
|
|
34 |
num_inference_steps: int = 30,
|
35 |
guidance_scale: float = 5.0,
|
36 |
adapter_conditioning_scale: float = 1.0,
|
@@ -43,6 +44,7 @@ def run(
|
|
43 |
image=image,
|
44 |
prompt=prompt,
|
45 |
negative_prompt=negative_prompt,
|
|
|
46 |
num_inference_steps=num_inference_steps,
|
47 |
guidance_scale=guidance_scale,
|
48 |
adapter_conditioning_scale=adapter_conditioning_scale,
|
@@ -116,6 +118,7 @@ with gr.Blocks(css="style.css") as demo:
|
|
116 |
image,
|
117 |
prompt,
|
118 |
negative_prompt,
|
|
|
119 |
num_inference_steps,
|
120 |
guidance_scale,
|
121 |
adapter_conditioning_scale,
|
@@ -130,10 +133,6 @@ with gr.Blocks(css="style.css") as demo:
|
|
130 |
queue=False,
|
131 |
api_name=False,
|
132 |
).then(
|
133 |
-
fn=model.change_adapter,
|
134 |
-
inputs=adapter_name,
|
135 |
-
api_name=False,
|
136 |
-
).success(
|
137 |
fn=run,
|
138 |
inputs=inputs,
|
139 |
outputs=result,
|
@@ -146,10 +145,6 @@ with gr.Blocks(css="style.css") as demo:
|
|
146 |
queue=False,
|
147 |
api_name=False,
|
148 |
).then(
|
149 |
-
fn=model.change_adapter,
|
150 |
-
inputs=adapter_name,
|
151 |
-
api_name=False,
|
152 |
-
).success(
|
153 |
fn=run,
|
154 |
inputs=inputs,
|
155 |
outputs=result,
|
@@ -162,10 +157,6 @@ with gr.Blocks(css="style.css") as demo:
|
|
162 |
queue=False,
|
163 |
api_name=False,
|
164 |
).then(
|
165 |
-
fn=model.change_adapter,
|
166 |
-
inputs=adapter_name,
|
167 |
-
api_name=False,
|
168 |
-
).success(
|
169 |
fn=run,
|
170 |
inputs=inputs,
|
171 |
outputs=result,
|
|
|
31 |
image: PIL.Image.Image,
|
32 |
prompt: str,
|
33 |
negative_prompt: str,
|
34 |
+
adapter_name: str,
|
35 |
num_inference_steps: int = 30,
|
36 |
guidance_scale: float = 5.0,
|
37 |
adapter_conditioning_scale: float = 1.0,
|
|
|
44 |
image=image,
|
45 |
prompt=prompt,
|
46 |
negative_prompt=negative_prompt,
|
47 |
+
adapter_name=adapter_name,
|
48 |
num_inference_steps=num_inference_steps,
|
49 |
guidance_scale=guidance_scale,
|
50 |
adapter_conditioning_scale=adapter_conditioning_scale,
|
|
|
118 |
image,
|
119 |
prompt,
|
120 |
negative_prompt,
|
121 |
+
adapter_name,
|
122 |
num_inference_steps,
|
123 |
guidance_scale,
|
124 |
adapter_conditioning_scale,
|
|
|
133 |
queue=False,
|
134 |
api_name=False,
|
135 |
).then(
|
|
|
|
|
|
|
|
|
136 |
fn=run,
|
137 |
inputs=inputs,
|
138 |
outputs=result,
|
|
|
145 |
queue=False,
|
146 |
api_name=False,
|
147 |
).then(
|
|
|
|
|
|
|
|
|
148 |
fn=run,
|
149 |
inputs=inputs,
|
150 |
outputs=result,
|
|
|
157 |
queue=False,
|
158 |
api_name=False,
|
159 |
).then(
|
|
|
|
|
|
|
|
|
160 |
fn=run,
|
161 |
inputs=inputs,
|
162 |
outputs=result,
|
model.py
CHANGED
@@ -1,4 +1,6 @@
|
|
1 |
-
|
|
|
|
|
2 |
|
3 |
import PIL.Image
|
4 |
import torch
|
@@ -26,74 +28,149 @@ ADAPTER_NAMES = [
|
|
26 |
]
|
27 |
|
28 |
|
29 |
-
class
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
def __init__(self):
|
31 |
self.model = CannyDetector()
|
32 |
|
|
|
|
|
|
|
33 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
34 |
return self.model(image, detect_resolution=384, image_resolution=1024)
|
35 |
|
36 |
|
37 |
-
class LineartPreprocessor:
|
38 |
def __init__(self):
|
39 |
-
|
40 |
-
|
|
|
|
|
41 |
|
42 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
43 |
return self.model(image, detect_resolution=384, image_resolution=1024)
|
44 |
|
45 |
|
46 |
-
class MidasPreprocessor:
|
47 |
def __init__(self):
|
48 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
49 |
self.model = MidasDetector.from_pretrained(
|
50 |
"valhalla/t2iadapter-aux-models", filename="dpt_large_384.pt", model_type="dpt_large"
|
51 |
-
)
|
|
|
|
|
|
|
52 |
|
53 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
54 |
return self.model(image, detect_resolution=512, image_resolution=1024)
|
55 |
|
56 |
|
57 |
-
class PidiNetPreprocessor:
|
58 |
def __init__(self):
|
59 |
-
|
60 |
-
|
|
|
|
|
61 |
|
62 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
63 |
return self.model(image, detect_resolution=512, image_resolution=1024, apply_filter=True)
|
64 |
|
65 |
|
66 |
-
class RecolorPreprocessor:
|
|
|
|
|
|
|
67 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
68 |
return image.convert("L").convert("RGB")
|
69 |
|
70 |
|
71 |
-
class ZoePreprocessor:
|
72 |
def __init__(self):
|
73 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
74 |
self.model = ZoeDetector.from_pretrained(
|
75 |
"valhalla/t2iadapter-aux-models", filename="zoed_nk.pth", model_type="zoedepth_nk"
|
76 |
-
)
|
|
|
|
|
|
|
77 |
|
78 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
79 |
return self.model(image, gamma_corrected=True, image_resolution=1024)
|
80 |
|
81 |
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
|
98 |
|
99 |
class Model:
|
@@ -103,11 +180,12 @@ class Model:
|
|
103 |
if adapter_name not in ADAPTER_NAMES:
|
104 |
raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")
|
105 |
|
|
|
106 |
self.adapter_name = adapter_name
|
107 |
|
108 |
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
109 |
if torch.cuda.is_available():
|
110 |
-
self.preprocessor = get_preprocessor(adapter_name)
|
111 |
|
112 |
model_id = "stabilityai/stable-diffusion-xl-base-1.0"
|
113 |
adapter = T2IAdapter.from_pretrained(
|
@@ -127,27 +205,39 @@ class Model:
|
|
127 |
).to(self.device)
|
128 |
self.pipe.enable_xformers_memory_efficient_attention()
|
129 |
else:
|
|
|
130 |
self.pipe = None
|
131 |
|
132 |
-
def
|
133 |
-
if not torch.cuda.is_available():
|
134 |
-
raise RuntimeError("This demo does not work on CPU.")
|
135 |
if adapter_name not in ADAPTER_NAMES:
|
136 |
raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")
|
137 |
-
if adapter_name == self.
|
138 |
return
|
139 |
|
140 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
141 |
torch.cuda.empty_cache()
|
142 |
-
self.preprocessor = get_preprocessor(adapter_name)
|
143 |
|
144 |
-
|
145 |
-
|
|
|
|
|
|
|
146 |
self.pipe.adapter = T2IAdapter.from_pretrained(
|
147 |
adapter_name,
|
148 |
torch_dtype=torch.float16,
|
149 |
varient="fp16",
|
150 |
).to(self.device)
|
|
|
|
|
|
|
151 |
|
152 |
def resize_image(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
153 |
w, h = image.size
|
@@ -161,6 +251,7 @@ class Model:
|
|
161 |
image: PIL.Image.Image,
|
162 |
prompt: str,
|
163 |
negative_prompt: str,
|
|
|
164 |
num_inference_steps: int = 30,
|
165 |
guidance_scale: float = 5.0,
|
166 |
adapter_conditioning_scale: float = 1.0,
|
@@ -168,12 +259,17 @@ class Model:
|
|
168 |
seed: int = 0,
|
169 |
apply_preprocess: bool = True,
|
170 |
) -> list[PIL.Image.Image]:
|
|
|
|
|
171 |
if num_inference_steps > self.MAX_NUM_INFERENCE_STEPS:
|
172 |
raise ValueError(f"Number of steps must be less than {self.MAX_NUM_INFERENCE_STEPS}")
|
173 |
|
174 |
# Resize image to avoid OOM
|
175 |
image = self.resize_image(image)
|
176 |
|
|
|
|
|
|
|
177 |
if apply_preprocess:
|
178 |
image = self.preprocessor(image)
|
179 |
|
|
|
1 |
+
import gc
|
2 |
+
import os
|
3 |
+
from abc import ABC, abstractmethod
|
4 |
|
5 |
import PIL.Image
|
6 |
import torch
|
|
|
28 |
]
|
29 |
|
30 |
|
31 |
+
class Preprocessor(ABC):
|
32 |
+
@abstractmethod
|
33 |
+
def to(self, device: torch.device | str) -> "Preprocessor":
|
34 |
+
pass
|
35 |
+
|
36 |
+
@abstractmethod
|
37 |
+
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
38 |
+
pass
|
39 |
+
|
40 |
+
|
41 |
+
class CannyPreprocessor(Preprocessor):
|
42 |
def __init__(self):
|
43 |
self.model = CannyDetector()
|
44 |
|
45 |
+
def to(self, device: torch.device | str) -> Preprocessor:
|
46 |
+
return self
|
47 |
+
|
48 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
49 |
return self.model(image, detect_resolution=384, image_resolution=1024)
|
50 |
|
51 |
|
52 |
+
class LineartPreprocessor(Preprocessor):
|
53 |
def __init__(self):
|
54 |
+
self.model = LineartDetector.from_pretrained("lllyasviel/Annotators")
|
55 |
+
|
56 |
+
def to(self, device: torch.device | str) -> Preprocessor:
|
57 |
+
return self.model.to(device)
|
58 |
|
59 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
60 |
return self.model(image, detect_resolution=384, image_resolution=1024)
|
61 |
|
62 |
|
63 |
+
class MidasPreprocessor(Preprocessor):
|
64 |
def __init__(self):
|
|
|
65 |
self.model = MidasDetector.from_pretrained(
|
66 |
"valhalla/t2iadapter-aux-models", filename="dpt_large_384.pt", model_type="dpt_large"
|
67 |
+
)
|
68 |
+
|
69 |
+
def to(self, device: torch.device | str) -> Preprocessor:
|
70 |
+
return self.model.to(device)
|
71 |
|
72 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
73 |
return self.model(image, detect_resolution=512, image_resolution=1024)
|
74 |
|
75 |
|
76 |
+
class PidiNetPreprocessor(Preprocessor):
|
77 |
def __init__(self):
|
78 |
+
self.model = PidiNetDetector.from_pretrained("lllyasviel/Annotators")
|
79 |
+
|
80 |
+
def to(self, device: torch.device | str) -> Preprocessor:
|
81 |
+
return self.model.to(device)
|
82 |
|
83 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
84 |
return self.model(image, detect_resolution=512, image_resolution=1024, apply_filter=True)
|
85 |
|
86 |
|
87 |
+
class RecolorPreprocessor(Preprocessor):
|
88 |
+
def to(self, device: torch.device | str) -> Preprocessor:
|
89 |
+
return self
|
90 |
+
|
91 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
92 |
return image.convert("L").convert("RGB")
|
93 |
|
94 |
|
95 |
+
class ZoePreprocessor(Preprocessor):
|
96 |
def __init__(self):
|
|
|
97 |
self.model = ZoeDetector.from_pretrained(
|
98 |
"valhalla/t2iadapter-aux-models", filename="zoed_nk.pth", model_type="zoedepth_nk"
|
99 |
+
)
|
100 |
+
|
101 |
+
def to(self, device: torch.device | str) -> Preprocessor:
|
102 |
+
return self.model.to(device)
|
103 |
|
104 |
def __call__(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
105 |
return self.model(image, gamma_corrected=True, image_resolution=1024)
|
106 |
|
107 |
|
108 |
+
PRELOAD_PREPROCESSORS_IN_GPU_MEMORY = os.getenv("PRELOAD_PREPROCESSORS_IN_GPU_MEMORY", "1") == "1"
|
109 |
+
PRELOAD_PREPROCESSORS_IN_CPU_MEMORY = os.getenv("PRELOAD_PREPROCESSORS_IN_CPU_MEMORY", "0") == "1"
|
110 |
+
if PRELOAD_PREPROCESSORS_IN_GPU_MEMORY:
|
111 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
112 |
+
preprocessors_gpu: dict[str, Preprocessor] = {
|
113 |
+
"TencentARC/t2i-adapter-canny-sdxl-1.0": CannyPreprocessor().to(device),
|
114 |
+
"TencentARC/t2i-adapter-sketch-sdxl-1.0": PidiNetPreprocessor().to(device),
|
115 |
+
"TencentARC/t2i-adapter-lineart-sdxl-1.0": LineartPreprocessor().to(device),
|
116 |
+
"TencentARC/t2i-adapter-depth-midas-sdxl-1.0": MidasPreprocessor().to(device),
|
117 |
+
"TencentARC/t2i-adapter-depth-zoe-sdxl-1.0": ZoePreprocessor().to(device),
|
118 |
+
"TencentARC/t2i-adapter-recolor-sdxl-1.0": RecolorPreprocessor().to(device),
|
119 |
+
}
|
120 |
+
|
121 |
+
def get_preprocessor(adapter_name: str) -> Preprocessor:
|
122 |
+
return preprocessors_gpu[adapter_name]
|
123 |
+
|
124 |
+
elif PRELOAD_PREPROCESSORS_IN_CPU_MEMORY:
|
125 |
+
preprocessors_cpu: dict[str, Preprocessor] = {
|
126 |
+
"TencentARC/t2i-adapter-canny-sdxl-1.0": CannyPreprocessor(),
|
127 |
+
"TencentARC/t2i-adapter-sketch-sdxl-1.0": PidiNetPreprocessor(),
|
128 |
+
"TencentARC/t2i-adapter-lineart-sdxl-1.0": LineartPreprocessor(),
|
129 |
+
"TencentARC/t2i-adapter-depth-midas-sdxl-1.0": MidasPreprocessor(),
|
130 |
+
"TencentARC/t2i-adapter-depth-zoe-sdxl-1.0": ZoePreprocessor(),
|
131 |
+
"TencentARC/t2i-adapter-recolor-sdxl-1.0": RecolorPreprocessor(),
|
132 |
+
}
|
133 |
+
|
134 |
+
def get_preprocessor(adapter_name: str) -> Preprocessor:
|
135 |
+
return preprocessors_cpu[adapter_name]
|
136 |
+
|
137 |
+
else:
|
138 |
+
|
139 |
+
def get_preprocessor(adapter_name: str) -> Preprocessor:
|
140 |
+
if adapter_name == "TencentARC/t2i-adapter-canny-sdxl-1.0":
|
141 |
+
return CannyPreprocessor()
|
142 |
+
elif adapter_name == "TencentARC/t2i-adapter-sketch-sdxl-1.0":
|
143 |
+
return PidiNetPreprocessor()
|
144 |
+
elif adapter_name == "TencentARC/t2i-adapter-lineart-sdxl-1.0":
|
145 |
+
return LineartPreprocessor()
|
146 |
+
elif adapter_name == "TencentARC/t2i-adapter-depth-midas-sdxl-1.0":
|
147 |
+
return MidasPreprocessor()
|
148 |
+
elif adapter_name == "TencentARC/t2i-adapter-depth-zoe-sdxl-1.0":
|
149 |
+
return ZoePreprocessor()
|
150 |
+
elif adapter_name == "TencentARC/t2i-adapter-recolor-sdxl-1.0":
|
151 |
+
return RecolorPreprocessor()
|
152 |
+
else:
|
153 |
+
raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")
|
154 |
+
|
155 |
+
def download_all_preprocessors():
|
156 |
+
for adapter_name in ADAPTER_NAMES:
|
157 |
+
get_preprocessor(adapter_name)
|
158 |
+
gc.collect()
|
159 |
+
|
160 |
+
download_all_preprocessors()
|
161 |
+
|
162 |
+
|
163 |
+
def download_all_adapters():
|
164 |
+
for adapter_name in ADAPTER_NAMES:
|
165 |
+
T2IAdapter.from_pretrained(
|
166 |
+
adapter_name,
|
167 |
+
torch_dtype=torch.float16,
|
168 |
+
varient="fp16",
|
169 |
+
)
|
170 |
+
gc.collect()
|
171 |
+
|
172 |
+
|
173 |
+
download_all_adapters()
|
174 |
|
175 |
|
176 |
class Model:
|
|
|
180 |
if adapter_name not in ADAPTER_NAMES:
|
181 |
raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")
|
182 |
|
183 |
+
self.preprocessor_name = adapter_name
|
184 |
self.adapter_name = adapter_name
|
185 |
|
186 |
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
187 |
if torch.cuda.is_available():
|
188 |
+
self.preprocessor = get_preprocessor(adapter_name).to(self.device)
|
189 |
|
190 |
model_id = "stabilityai/stable-diffusion-xl-base-1.0"
|
191 |
adapter = T2IAdapter.from_pretrained(
|
|
|
205 |
).to(self.device)
|
206 |
self.pipe.enable_xformers_memory_efficient_attention()
|
207 |
else:
|
208 |
+
self.preprocessor = None # type: ignore
|
209 |
self.pipe = None
|
210 |
|
211 |
+
def change_preprocessor(self, adapter_name: str) -> None:
|
|
|
|
|
212 |
if adapter_name not in ADAPTER_NAMES:
|
213 |
raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")
|
214 |
+
if adapter_name == self.preprocessor_name:
|
215 |
return
|
216 |
|
217 |
+
if PRELOAD_PREPROCESSORS_IN_GPU_MEMORY:
|
218 |
+
pass
|
219 |
+
elif PRELOAD_PREPROCESSORS_IN_CPU_MEMORY:
|
220 |
+
self.preprocessor.to("cpu")
|
221 |
+
else:
|
222 |
+
del self.preprocessor
|
223 |
+
self.preprocessor = get_preprocessor(adapter_name).to(self.device)
|
224 |
+
self.preprocessor_name = adapter_name
|
225 |
+
gc.collect()
|
226 |
torch.cuda.empty_cache()
|
|
|
227 |
|
228 |
+
def change_adapter(self, adapter_name: str) -> None:
|
229 |
+
if adapter_name not in ADAPTER_NAMES:
|
230 |
+
raise ValueError(f"Adapter name must be one of {ADAPTER_NAMES}")
|
231 |
+
if adapter_name == self.adapter_name:
|
232 |
+
return
|
233 |
self.pipe.adapter = T2IAdapter.from_pretrained(
|
234 |
adapter_name,
|
235 |
torch_dtype=torch.float16,
|
236 |
varient="fp16",
|
237 |
).to(self.device)
|
238 |
+
self.adapter_name = adapter_name
|
239 |
+
gc.collect()
|
240 |
+
torch.cuda.empty_cache()
|
241 |
|
242 |
def resize_image(self, image: PIL.Image.Image) -> PIL.Image.Image:
|
243 |
w, h = image.size
|
|
|
251 |
image: PIL.Image.Image,
|
252 |
prompt: str,
|
253 |
negative_prompt: str,
|
254 |
+
adapter_name: str,
|
255 |
num_inference_steps: int = 30,
|
256 |
guidance_scale: float = 5.0,
|
257 |
adapter_conditioning_scale: float = 1.0,
|
|
|
259 |
seed: int = 0,
|
260 |
apply_preprocess: bool = True,
|
261 |
) -> list[PIL.Image.Image]:
|
262 |
+
if not torch.cuda.is_available():
|
263 |
+
raise RuntimeError("This demo does not work on CPU.")
|
264 |
if num_inference_steps > self.MAX_NUM_INFERENCE_STEPS:
|
265 |
raise ValueError(f"Number of steps must be less than {self.MAX_NUM_INFERENCE_STEPS}")
|
266 |
|
267 |
# Resize image to avoid OOM
|
268 |
image = self.resize_image(image)
|
269 |
|
270 |
+
self.change_preprocessor(adapter_name)
|
271 |
+
self.change_adapter(adapter_name)
|
272 |
+
|
273 |
if apply_preprocess:
|
274 |
image = self.preprocessor(image)
|
275 |
|