Linaqruf commited on
Commit
158fb03
1 Parent(s): 16f7e05

update demo

Browse files
.gitattributes CHANGED
@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ images/amelia-watson.png filter=lfs diff=lfs merge=lfs -text
37
+ images/furina.png filter=lfs diff=lfs merge=lfs -text
38
+ images/pastel-style.png filter=lfs diff=lfs merge=lfs -text
39
+ images/ufotable-style.png filter=lfs diff=lfs merge=lfs -text
app.py CHANGED
@@ -4,46 +4,52 @@ from __future__ import annotations
4
 
5
  import os
6
  import random
7
-
8
  import gradio as gr
9
  import numpy as np
10
  import PIL.Image
11
  import torch
 
 
 
 
 
 
12
  from diffusers.models import AutoencoderKL
13
- from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler
14
 
15
- DESCRIPTION = '# Animagine XL'
16
  if not torch.cuda.is_available():
17
- DESCRIPTION += '\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>'
18
-
19
  MAX_SEED = np.iinfo(np.int32).max
20
- CACHE_EXAMPLES = torch.cuda.is_available() and os.getenv(
21
- 'CACHE_EXAMPLES') == '1'
22
- MAX_IMAGE_SIZE = int(os.getenv('MAX_IMAGE_SIZE', '2048'))
23
- USE_TORCH_COMPILE = os.getenv('USE_TORCH_COMPILE') == '1'
24
- ENABLE_CPU_OFFLOAD = os.getenv('ENABLE_CPU_OFFLOAD') == '1'
25
 
26
  MODEL = "Linaqruf/animagine-xl"
27
 
28
- device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
29
  if torch.cuda.is_available():
30
- pipe = StableDiffusionXLPipeline.from_pretrained(
31
  MODEL,
32
  torch_dtype=torch.float16,
 
33
  use_safetensors=True,
34
- variant='fp16')
 
35
 
36
  pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
37
-
38
  if ENABLE_CPU_OFFLOAD:
39
  pipe.enable_model_cpu_offload()
40
  else:
41
  pipe.to(device)
42
 
43
  if USE_TORCH_COMPILE:
44
- pipe.unet = torch.compile(pipe.unet,
45
- mode='reduce-overhead',
46
- fullgraph=True)
47
  else:
48
  pipe = None
49
 
@@ -54,197 +60,423 @@ def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
54
  return seed
55
 
56
 
57
- def generate(prompt: str,
58
- negative_prompt: str = '',
59
- prompt_2: str = '',
60
- negative_prompt_2: str = '',
61
- use_prompt_2: bool = False,
62
- seed: int = 0,
63
- width: int = 1024,
64
- height: int = 1024,
65
- target_width: int = 1024,
66
- target_height: int = 1024,
67
- original_width: int = 4096,
68
- original_height: int = 4096,
69
- guidance_scale_base: float = 12.0,
70
- num_inference_steps_base: int = 50) -> PIL.Image.Image:
71
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
72
  generator = torch.Generator().manual_seed(seed)
73
 
74
- if negative_prompt == '':
75
- negative_prompt = None # type: ignore
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
76
  if not use_prompt_2:
77
- prompt_2 = None # type: ignore
78
- negative_prompt_2 = None # type: ignore
79
- if negative_prompt_2 == '':
80
- negative_prompt_2 = None
81
-
82
- return pipe(prompt=prompt,
83
- negative_prompt=negative_prompt,
84
- prompt_2=prompt_2,
85
- negative_prompt_2=negative_prompt_2,
86
- width=width,
87
- height=height,
88
- target_size=(target_width, target_height),
89
- original_size=(original_width, original_height),
90
- guidance_scale=guidance_scale_base,
91
- num_inference_steps=num_inference_steps_base,
92
- generator=generator,
93
- output_type='pil').images[0]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94
 
95
 
96
  examples = [
97
- 'face focus, cute, masterpiece, best quality, 1girl, green hair, sweater, looking at viewer, upper body, beanie, outdoors, night, turtleneck',
98
- 'face focus, bishounen, masterpiece, best quality, 1boy, green hair, sweater, looking at viewer, upper body, beanie, outdoors, night, turtleneck',
99
  ]
100
 
101
- # choices = [
102
- # "Vertical (9:16)",
103
- # "Portrait (4:5)",
104
- # "Square (1:1)",
105
- # "Photo (4:3)",
106
- # "Landscape (3:2)",
107
- # "Widescreen (16:9)",
108
- # "Cinematic (21:9)",
109
- # ]
110
-
111
- # choice_to_size = {
112
- # "Vertical (9:16)": (768, 1344),
113
- # "Portrait (4:5)": (912, 1144),
114
- # "Square (1:1)": (1024, 1024),
115
- # "Photo (4:3)": (1184, 888),
116
- # "Landscape (3:2)": (1256, 832),
117
- # "Widescreen (16:9)": (1368, 768),
118
- # "Cinematic (21:9)": (1568, 672),
119
- # }
120
-
121
- with gr.Blocks(css='style.css', theme='NoCrypt/[email protected]') as demo:
122
- gr.Markdown(DESCRIPTION)
123
- gr.DuplicateButton(value='Duplicate Space for private use',
124
- elem_id='duplicate-button',
125
- visible=os.getenv('SHOW_DUPLICATE_BUTTON') == '1')
 
 
 
 
 
 
 
 
 
 
 
 
 
126
  with gr.Row():
127
  with gr.Column(scale=1):
128
- prompt = gr.Text(
129
- label='Prompt',
130
- max_lines=1,
131
- placeholder='Enter your prompt',
132
- )
133
- negative_prompt = gr.Text(
134
- label='Negative Prompt',
135
- max_lines=1,
136
- placeholder='Enter a negative prompt',
137
- value='lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry',
138
- )
139
- use_prompt_2 = gr.Checkbox(
140
- label='Use prompt 2',
141
- value=False
142
- )
143
- prompt_2 = gr.Text(
144
- label='Prompt 2',
145
- max_lines=1,
146
- placeholder='Enter your prompt',
147
- visible=False,
148
- )
149
- negative_prompt_2 = gr.Text(
150
- label='Negative prompt 2',
151
- max_lines=1,
152
- placeholder='Enter a negative prompt',
153
- visible=False,
154
- )
 
 
 
 
 
 
 
 
155
 
156
- # with gr.Row():
157
- # aspect_ratio = gr.Dropdown(choices=choices, label="Aspect Ratio Preset", value=choices[2])
158
- with gr.Row():
159
- width = gr.Slider(
160
- label='Width',
161
- minimum=256,
162
- maximum=MAX_IMAGE_SIZE,
163
- step=32,
164
- value=1024,
 
 
 
 
 
165
  )
166
- height = gr.Slider(
167
- label='Height',
168
- minimum=256,
169
- maximum=MAX_IMAGE_SIZE,
170
- step=32,
171
- value=1024,
172
  )
173
- with gr.Accordion(label='Advanced Config', open=False):
174
- with gr.Accordion(label='Conditioning Resolution', open=False):
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
175
  with gr.Row():
176
- original_width = gr.Slider(
177
- label='Original Width',
178
- minimum=1024,
179
- maximum=4096,
180
- step=32,
181
- value=4096,
182
  )
183
- original_height = gr.Slider(
184
- label='Original Height',
185
- minimum=1024,
186
- maximum=4096,
187
- step=32,
188
- value=4096,
189
  )
190
- with gr.Row():
191
- target_width = gr.Slider(
192
- label='Target Width',
193
- minimum=1024,
194
- maximum=4096,
195
- step=32,
196
- value=1024,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
197
  )
198
- target_height = gr.Slider(
199
- label='Target Height',
200
- minimum=1024,
201
- maximum=4096,
202
- step=32,
203
- value=1024,
204
- )
205
- seed = gr.Slider(label='Seed',
206
- minimum=0,
207
- maximum=MAX_SEED,
208
- step=1,
209
- value=0)
210
-
211
- randomize_seed = gr.Checkbox(label='Randomize seed', value=True)
212
- with gr.Row():
213
- guidance_scale_base = gr.Slider(
214
- label='Guidance scale',
215
- minimum=1,
216
- maximum=20,
217
- step=0.1,
218
- value=12.0)
219
- num_inference_steps_base = gr.Slider(
220
- label='Number of inference steps',
221
- minimum=10,
222
- maximum=100,
223
- step=1,
224
- value=50)
225
-
226
  with gr.Column(scale=2):
227
  with gr.Blocks():
228
- run_button = gr.Button('Generate')
229
- result = gr.Image(label='Result', show_label=False)
230
-
231
- gr.Examples(examples=examples,
232
- inputs=prompt,
233
- outputs=result,
234
- fn=generate,
235
- cache_examples=CACHE_EXAMPLES)
236
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
237
  use_prompt_2.change(
238
  fn=lambda x: gr.update(visible=x),
239
  inputs=use_prompt_2,
240
- outputs=prompt_2,
241
  queue=False,
242
  api_name=False,
243
  )
244
- use_prompt_2.change(
 
 
 
 
 
 
 
245
  fn=lambda x: gr.update(visible=x),
246
- inputs=use_prompt_2,
247
- outputs=negative_prompt_2,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
248
  queue=False,
249
  api_name=False,
250
  )
@@ -262,8 +494,13 @@ with gr.Blocks(css='style.css', theme='NoCrypt/[email protected]') as demo:
262
  target_height,
263
  original_width,
264
  original_height,
265
- guidance_scale_base,
266
- num_inference_steps_base,
 
 
 
 
 
267
  ]
268
  prompt.submit(
269
  fn=randomize_seed_fn,
@@ -275,7 +512,7 @@ with gr.Blocks(css='style.css', theme='NoCrypt/[email protected]') as demo:
275
  fn=generate,
276
  inputs=inputs,
277
  outputs=result,
278
- api_name='run',
279
  )
280
  negative_prompt.submit(
281
  fn=randomize_seed_fn,
@@ -326,4 +563,4 @@ with gr.Blocks(css='style.css', theme='NoCrypt/[email protected]') as demo:
326
  api_name=False,
327
  )
328
 
329
- demo.queue(max_size=20).launch()
 
4
 
5
  import os
6
  import random
7
+ import toml
8
  import gradio as gr
9
  import numpy as np
10
  import PIL.Image
11
  import torch
12
+ import utils
13
+ import gc
14
+ from safetensors.torch import load_file
15
+ import lora_diffusers
16
+ from lora_diffusers import LoRANetwork, create_network_from_weights
17
+ from huggingface_hub import hf_hub_download
18
  from diffusers.models import AutoencoderKL
19
+ from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler
20
 
21
+ DESCRIPTION = "Animagine XL"
22
  if not torch.cuda.is_available():
23
+ DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"
24
+ IS_COLAB = utils.is_google_colab()
25
  MAX_SEED = np.iinfo(np.int32).max
26
+ CACHE_EXAMPLES = torch.cuda.is_available() and os.getenv("CACHE_EXAMPLES") == "1"
27
+ MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "2048"))
28
+ USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE") == "1"
29
+ ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD") == "1"
 
30
 
31
  MODEL = "Linaqruf/animagine-xl"
32
 
33
+ device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
34
  if torch.cuda.is_available():
35
+ pipe = DiffusionPipeline.from_pretrained(
36
  MODEL,
37
  torch_dtype=torch.float16,
38
+ custom_pipeline="lpw_stable_diffusion_xl.py",
39
  use_safetensors=True,
40
+ variant="fp16",
41
+ )
42
 
43
  pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
44
+
45
  if ENABLE_CPU_OFFLOAD:
46
  pipe.enable_model_cpu_offload()
47
  else:
48
  pipe.to(device)
49
 
50
  if USE_TORCH_COMPILE:
51
+ pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
52
+
 
53
  else:
54
  pipe = None
55
 
 
60
  return seed
61
 
62
 
63
+ def get_image_path(base_path):
64
+ extensions = [".jpg", ".jpeg", ".png", ".bmp", ".gif"]
65
+ for ext in extensions:
66
+ if os.path.exists(base_path + ext):
67
+ return base_path + ext
68
+ # If no match is found, return None or raise an error
69
+ return None
70
+
71
+
72
+ def update_selection(selected_state: gr.SelectData):
73
+ lora_repo = sdxl_loras[selected_state.index]["repo"]
74
+ lora_weight = sdxl_loras[selected_state.index]["multiplier"]
75
+ updated_selected_info = f"{lora_repo}"
76
+ updated_prompt = sdxl_loras[selected_state.index]["sample_prompt"]
77
+ updated_negative = sdxl_loras[selected_state.index]["sample_negative"]
78
+
79
+ return (
80
+ updated_selected_info,
81
+ selected_state,
82
+ lora_weight,
83
+ updated_prompt,
84
+ negative_presets_dict.get(updated_negative, ""),
85
+ updated_negative,
86
+ )
87
+
88
+
89
+ def create_network(text_encoders, unet, state_dict, multiplier, device):
90
+ network = create_network_from_weights(
91
+ text_encoders, unet, state_dict, multiplier=multiplier
92
+ )
93
+ network.load_state_dict(state_dict)
94
+ network.to(device, dtype=unet.dtype)
95
+ network.apply_to(multiplier=multiplier)
96
+ return network
97
+
98
+
99
+ # def backup_sd(state_dict):
100
+ # for k, v in state_dict.items():
101
+ # state_dict[k] = v.detach().cpu()
102
+ # return state_dict
103
+
104
+
105
+ def generate(
106
+ prompt: str,
107
+ negative_prompt: str = "",
108
+ prompt_2: str = "",
109
+ negative_prompt_2: str = "",
110
+ use_prompt_2: bool = False,
111
+ seed: int = 0,
112
+ width: int = 1024,
113
+ height: int = 1024,
114
+ target_width: int = 1024,
115
+ target_height: int = 1024,
116
+ original_width: int = 4096,
117
+ original_height: int = 4096,
118
+ guidance_scale: float = 12.0,
119
+ num_inference_steps: int = 50,
120
+ use_lora: bool = False,
121
+ lora_weight: float = 1.0,
122
+ set_target_size: bool = False,
123
+ set_original_size: bool = False,
124
+ selected_state: str = "",
125
+ ) -> PIL.Image.Image:
126
  generator = torch.Generator().manual_seed(seed)
127
 
128
+ network = None # Initialize to None
129
+ network_state = {"current_lora": None, "multiplier": None}
130
+
131
+ # _unet = pipe.unet.state_dict()
132
+ # backup_sd(_unet)
133
+ # _text_encoder = pipe.text_encoder.state_dict()
134
+ # backup_sd(_text_encoder)
135
+ # _text_encoder_2 = pipe.text_encoder_2.state_dict()
136
+ # backup_sd(_text_encoder_2)
137
+
138
+ if not set_original_size:
139
+ original_width = 4096
140
+ original_height = 4096
141
+ if not set_target_size:
142
+ target_width = width
143
+ target_height = height
144
+ if negative_prompt == "":
145
+ negative_prompt = None
146
  if not use_prompt_2:
147
+ prompt_2 = None
148
+ negative_prompt_2 = None
149
+ if negative_prompt_2 == "":
150
+ negative_prompt_2 = None
151
+
152
+ if use_lora:
153
+ if not selected_state:
154
+ raise Exception("You must select a LoRA")
155
+
156
+ repo_name = sdxl_loras[selected_state.index]["repo"]
157
+ full_path_lora = saved_names[selected_state.index]
158
+ weight_name = sdxl_loras[selected_state.index]["weights"]
159
+
160
+ lora_sd = load_file(full_path_lora)
161
+ text_encoders = [pipe.text_encoder, pipe.text_encoder_2]
162
+
163
+ if network_state["current_lora"] != repo_name:
164
+ network = create_network(
165
+ text_encoders, pipe.unet, lora_sd, lora_weight, device
166
+ )
167
+ network_state["current_lora"] = repo_name
168
+ network_state["multiplier"] = lora_weight
169
+
170
+ elif network_state["multiplier"] != lora_weight:
171
+ network = create_network(
172
+ text_encoders, pipe.unet, lora_sd, lora_weight, device
173
+ )
174
+ network_state["multiplier"] = lora_weight
175
+ else:
176
+ if network:
177
+ network.unapply_to()
178
+ network = None
179
+ network_state = {"current_lora": None, "multiplier": None}
180
+
181
+ try:
182
+ image = pipe(
183
+ prompt=prompt,
184
+ negative_prompt=negative_prompt,
185
+ prompt_2=prompt_2,
186
+ negative_prompt_2=negative_prompt_2,
187
+ width=width,
188
+ height=height,
189
+ target_size=(target_width, target_height),
190
+ original_size=(original_width, original_height),
191
+ guidance_scale=guidance_scale,
192
+ num_inference_steps=num_inference_steps,
193
+ generator=generator,
194
+ output_type="pil",
195
+ ).images[0]
196
+
197
+ if network:
198
+ network.unapply_to()
199
+ network = None
200
+
201
+ return image
202
+
203
+ except Exception as e:
204
+ print(f"An error occurred: {e}")
205
+ raise
206
+
207
+ finally:
208
+ # pipe.unet.load_state_dict(_unet)
209
+ # pipe.text_encoder.load_state_dict(_text_encoder)
210
+ # pipe.text_encoder_2.load_state_dict(_text_encoder_2)
211
+
212
+ # del _unet, _text_encoder, _text_encoder_2
213
+
214
+ if network:
215
+ network.unapply_to()
216
+ network = None
217
+
218
+ if use_lora:
219
+ del lora_sd, text_encoders
220
+ gc.collect()
221
 
222
 
223
  examples = [
224
+ "face focus, cute, masterpiece, best quality, 1girl, green hair, sweater, looking at viewer, upper body, beanie, outdoors, night, turtleneck",
225
+ "face focus, bishounen, masterpiece, best quality, 1boy, green hair, sweater, looking at viewer, upper body, beanie, outdoors, night, turtleneck",
226
  ]
227
 
228
+ negative_presets_dict = {
229
+ "None": "",
230
+ "Standard": "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry",
231
+ "Weighted": "(low quality, worst quality:1.2), 3d, watermark, signature, ugly, poorly drawn, bad image",
232
+ }
233
+
234
+ with open("lora.toml", "r") as file:
235
+ data = toml.load(file)
236
+ sdxl_loras = [
237
+ {
238
+ "image": get_image_path(item["image"]),
239
+ "title": item["title"],
240
+ "repo": item["repo"],
241
+ "weights": item["weights"],
242
+ "multiplier": item["multiplier"] if "multiplier" in item else "1.0",
243
+ "sample_prompt": item["sample_prompt"],
244
+ "sample_negative": item["sample_negative"],
245
+ }
246
+ for item in data["data"]
247
+ ]
248
+ saved_names = [hf_hub_download(item["repo"], item["weights"]) for item in sdxl_loras]
249
+
250
+
251
+ with gr.Blocks(css="style.css", theme="NoCrypt/[email protected]") as demo:
252
+ title = gr.HTML(
253
+ f"""<h1><span>{DESCRIPTION}</span></h1>""",
254
+ elem_id="title",
255
+ )
256
+ gr.Markdown(
257
+ f"""Gradio demo for [Linaqruf/animagine-xl](https://huggingface.co/spaces/Linaqruf/Animagine-XL)""",
258
+ elem_id="subtitle",
259
+ )
260
+ gr.DuplicateButton(
261
+ value="Duplicate Space for private use",
262
+ elem_id="duplicate-button",
263
+ visible=os.getenv("SHOW_DUPLICATE_BUTTON") == "1",
264
+ )
265
+ selected_state = gr.State()
266
  with gr.Row():
267
  with gr.Column(scale=1):
268
+ with gr.Group():
269
+ prompt = gr.Text(
270
+ label="Prompt",
271
+ max_lines=5,
272
+ placeholder="Enter your prompt",
273
+ )
274
+ negative_prompt = gr.Text(
275
+ label="Negative Prompt",
276
+ max_lines=5,
277
+ placeholder="Enter a negative prompt",
278
+ value="lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry",
279
+ )
280
+ with gr.Accordion(label="Negative Presets", open=False):
281
+ negative_presets = gr.Dropdown(
282
+ label="Negative Presets",
283
+ show_label=False,
284
+ choices=list(negative_presets_dict.keys()),
285
+ value="Standard",
286
+ )
287
+
288
+ with gr.Row():
289
+ use_prompt_2 = gr.Checkbox(label="Use prompt 2", value=False)
290
+ use_lora = gr.Checkbox(label="Use LoRA", value=False)
291
+
292
+ with gr.Group(visible=False) as prompt2_group:
293
+ prompt_2 = gr.Text(
294
+ label="Prompt 2",
295
+ max_lines=5,
296
+ placeholder="Enter your prompt",
297
+ )
298
+ negative_prompt_2 = gr.Text(
299
+ label="Negative prompt 2",
300
+ max_lines=5,
301
+ placeholder="Enter a negative prompt",
302
+ )
303
 
304
+ with gr.Group(visible=False) as lora_group:
305
+ selector_info = gr.Text(
306
+ label="Selected LoRA",
307
+ max_lines=1,
308
+ value="No LoRA selected.",
309
+ )
310
+ lora_selection = gr.Gallery(
311
+ value=[(item["image"], item["title"]) for item in sdxl_loras],
312
+ label="Animagine XL LoRA",
313
+ show_label=False,
314
+ allow_preview=False,
315
+ columns=2,
316
+ elem_id="gallery",
317
+ show_share_button=False,
318
  )
319
+ lora_weight = gr.Slider(
320
+ label="Multiplier",
321
+ minimum=0,
322
+ maximum=1,
323
+ step=0.05,
324
+ value=1,
325
  )
326
+
327
+ with gr.Group():
328
+ with gr.Row():
329
+ width = gr.Slider(
330
+ label="Width",
331
+ minimum=256,
332
+ maximum=MAX_IMAGE_SIZE,
333
+ step=32,
334
+ value=1024,
335
+ )
336
+ height = gr.Slider(
337
+ label="Height",
338
+ minimum=256,
339
+ maximum=MAX_IMAGE_SIZE,
340
+ step=32,
341
+ value=1024,
342
+ )
343
+
344
+ with gr.Accordion(label="Advanced Options", open=False):
345
+ seed = gr.Slider(
346
+ label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0
347
+ )
348
+
349
+ randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
350
+
351
  with gr.Row():
352
+ guidance_scale = gr.Slider(
353
+ label="Guidance scale",
354
+ minimum=1,
355
+ maximum=20,
356
+ step=0.1,
357
+ value=12.0,
358
  )
359
+ num_inference_steps = gr.Slider(
360
+ label="Number of inference steps",
361
+ minimum=10,
362
+ maximum=100,
363
+ step=1,
364
+ value=50,
365
  )
366
+ with gr.Group():
367
+ with gr.Row():
368
+ set_target_size = gr.Checkbox(
369
+ label="Target Size", value=False
370
+ )
371
+ set_original_size = gr.Checkbox(
372
+ label="Original Size", value=False
373
+ )
374
+ with gr.Group():
375
+ with gr.Row():
376
+ original_width = gr.Slider(
377
+ label="Original Width",
378
+ minimum=1024,
379
+ maximum=4096,
380
+ step=32,
381
+ value=4096,
382
+ visible=False,
383
+ )
384
+ original_height = gr.Slider(
385
+ label="Original Height",
386
+ minimum=1024,
387
+ maximum=4096,
388
+ step=32,
389
+ value=4096,
390
+ visible=False,
391
+ )
392
+ with gr.Row():
393
+ target_width = gr.Slider(
394
+ label="Target Width",
395
+ minimum=1024,
396
+ maximum=4096,
397
+ step=32,
398
+ value=width.value,
399
+ visible=False,
400
+ )
401
+ target_height = gr.Slider(
402
+ label="Target Height",
403
+ minimum=1024,
404
+ maximum=4096,
405
+ step=32,
406
+ value=height.value,
407
+ visible=False,
408
  )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
409
  with gr.Column(scale=2):
410
  with gr.Blocks():
411
+ run_button = gr.Button("Generate", variant="primary")
412
+ result = gr.Image(label="Result", show_label=False)
413
+
414
+ gr.Examples(
415
+ examples=examples,
416
+ inputs=prompt,
417
+ outputs=result,
418
+ fn=generate,
419
+ cache_examples=CACHE_EXAMPLES,
420
+ )
421
+ lora_selection.select(
422
+ update_selection,
423
+ outputs=[
424
+ selector_info,
425
+ selected_state,
426
+ lora_weight,
427
+ prompt,
428
+ negative_prompt,
429
+ negative_presets,
430
+ ],
431
+ queue=False,
432
+ show_progress=False,
433
+ )
434
  use_prompt_2.change(
435
  fn=lambda x: gr.update(visible=x),
436
  inputs=use_prompt_2,
437
+ outputs=prompt2_group,
438
  queue=False,
439
  api_name=False,
440
  )
441
+ negative_presets.change(
442
+ fn=lambda x: gr.update(value=negative_presets_dict.get(x, "")),
443
+ inputs=negative_presets,
444
+ outputs=negative_prompt,
445
+ queue=False,
446
+ api_name=False,
447
+ )
448
+ use_lora.change(
449
  fn=lambda x: gr.update(visible=x),
450
+ inputs=use_lora,
451
+ outputs=lora_group,
452
+ queue=False,
453
+ api_name=False,
454
+ )
455
+ set_target_size.change(
456
+ fn=lambda x: (gr.update(visible=x), gr.update(visible=x)),
457
+ inputs=set_target_size,
458
+ outputs=[target_width, target_height],
459
+ queue=False,
460
+ api_name=False,
461
+ )
462
+ set_original_size.change(
463
+ fn=lambda x: (gr.update(visible=x), gr.update(visible=x)),
464
+ inputs=set_original_size,
465
+ outputs=[original_width, original_height],
466
+ queue=False,
467
+ api_name=False,
468
+ )
469
+ width.change(
470
+ fn=lambda x: gr.update(value=x),
471
+ inputs=width,
472
+ outputs=target_width,
473
+ queue=False,
474
+ api_name=False,
475
+ )
476
+ height.change(
477
+ fn=lambda x: gr.update(value=x),
478
+ inputs=height,
479
+ outputs=target_height,
480
  queue=False,
481
  api_name=False,
482
  )
 
494
  target_height,
495
  original_width,
496
  original_height,
497
+ guidance_scale,
498
+ num_inference_steps,
499
+ use_lora,
500
+ lora_weight,
501
+ set_target_size,
502
+ set_original_size,
503
+ selected_state,
504
  ]
505
  prompt.submit(
506
  fn=randomize_seed_fn,
 
512
  fn=generate,
513
  inputs=inputs,
514
  outputs=result,
515
+ api_name="run",
516
  )
517
  negative_prompt.submit(
518
  fn=randomize_seed_fn,
 
563
  api_name=False,
564
  )
565
 
566
+ demo.queue(max_size=20).launch(debug=IS_COLAB, share=IS_COLAB)
demo.ipynb ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "cells": [
3
+ {
4
+ "cell_type": "code",
5
+ "execution_count": null,
6
+ "metadata": {
7
+ "id": "PeEyOhUDHhzF"
8
+ },
9
+ "outputs": [],
10
+ "source": [
11
+ "import os\n",
12
+ "import subprocess\n",
13
+ "\n",
14
+ "ROOT_DIR = \"/content\"\n",
15
+ "REPO_URL = \"https://huggingface.co/spaces/Linaqruf/Animagine-XL\"\n",
16
+ "REPO_DIR = os.path.join(ROOT_DIR, \"Animagine-XL\")\n",
17
+ "\n",
18
+ "def clone(url, dir, branch=None):\n",
19
+ " subprocess.run([\"git\", \"clone\", url, dir], check=True)\n",
20
+ " if branch:\n",
21
+ " subprocess.run([\"git\", \"checkout\", branch], cwd=dir, check=True)\n",
22
+ "\n",
23
+ "def install_deps(dir):\n",
24
+ " subprocess.run([\"pip\", \"install\", \"-r\", \"requirements.txt\"], cwd=dir, check=True)\n",
25
+ "\n",
26
+ "def main():\n",
27
+ " if not os.path.exists(REPO_DIR):\n",
28
+ " print(f\"Cloning Repository to {REPO_DIR}\")\n",
29
+ " clone(REPO_URL, REPO_DIR)\n",
30
+ " print(f\"Installing required python libraries\")\n",
31
+ " install_deps(REPO_DIR)\n",
32
+ " print(\"Done!\")\n",
33
+ "\n",
34
+ " os.chdir(REPO_DIR)\n",
35
+ " !python app.py\n",
36
+ "\n",
37
+ "if __name__ == \"__main__\":\n",
38
+ " main()\n"
39
+ ]
40
+ }
41
+ ],
42
+ "metadata": {
43
+ "accelerator": "GPU",
44
+ "colab": {
45
+ "machine_shape": "hm",
46
+ "provenance": [],
47
+ "gpuType": "A100"
48
+ },
49
+ "kernelspec": {
50
+ "display_name": "Python 3",
51
+ "name": "python3"
52
+ },
53
+ "language_info": {
54
+ "name": "python"
55
+ }
56
+ },
57
+ "nbformat": 4,
58
+ "nbformat_minor": 0
59
+ }
images/.placeholder ADDED
@@ -0,0 +1 @@
 
 
1
+
images/amelia-watson.png ADDED

Git LFS Details

  • SHA256: 83d6e7381fa4702608e4714ad349398efbc45ca6de0d20a0eb644fe5ecdcac34
  • Pointer size: 132 Bytes
  • Size of remote file: 1.68 MB
images/furina.png ADDED

Git LFS Details

  • SHA256: 290f2ae2e8cc132e64e43c19ee0c1ba5485259b28ad50eac36fc7146720f6575
  • Pointer size: 132 Bytes
  • Size of remote file: 1.61 MB
images/pastel-style.png ADDED

Git LFS Details

  • SHA256: d9f5bc5dd0f15d3e3f8a078a5bcb2ba375cb429817831c285b08f92a33d149f3
  • Pointer size: 132 Bytes
  • Size of remote file: 1.54 MB
images/ufotable-style.png ADDED

Git LFS Details

  • SHA256: 785eaf445be8c314d8978b7e2d909e52d49f5264c273ef98e094f86cdfa1a1e9
  • Pointer size: 132 Bytes
  • Size of remote file: 1.34 MB
lora.toml ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [[data]]
2
+ image = "images/pastel-style"
3
+ title = "Pastel Style"
4
+ repo = "Linaqruf/pastel-anime-xl-lora"
5
+ weights = "pastel-anime-xl-latest.safetensors"
6
+ multiplier = 0.6
7
+ sample_prompt = "face focus, cute, masterpiece, best quality, 1girl, green hair, sweater, looking at viewer, upper body, beanie, outdoors, night, turtleneck"
8
+ sample_negative = "Standard"
9
+
10
+ [[data]]
11
+ image = "images/ufotable-style"
12
+ title = "Ufotable Style"
13
+ repo = "Linaqruf/ufotable-xl-lora"
14
+ weights = "ufotable_style_xl.safetensors"
15
+ multiplier = 0.4
16
+ sample_prompt = "face focus, cute, masterpiece, best quality, bokeh, breasts, 1girl, solo, looking at viewer, long hair, white ribbon, smile, school uniform, bangs, black hair ribbon, swept bangs, sailor collar, serafuku, blush, ribbon, ahoge, brown eyes, long sleeves, collarbone, parted lips, sweater"
17
+ sample_negative = "Weighted"
18
+
19
+ [[data]]
20
+ image = "images/amelia-watson"
21
+ title = "Amelia Watson"
22
+ repo = "Linaqruf/amelia-watson-xl-lora"
23
+ weights = "amelia_watson_xl.safetensors"
24
+ multiplier = 0.5
25
+ sample_prompt = "face focus, masterpiece, best quality, amelia watson, bokeh, cute, 1girl, solo, monocle hair ornament, medium hair, brown eyewear, white shirt, red necktie, upper body, looking at viewer, blue eyes, leaf, plant"
26
+ sample_negative = "Weighted"
27
+
28
+ [[data]]
29
+ image = "images/furina"
30
+ title = "Furina"
31
+ repo = "Linaqruf/furina-xl-lora"
32
+ weights = "furina_xl.safetensors"
33
+ multiplier = 0.7
34
+ sample_prompt = "face focus, masterpiece, best quality, furina, bokeh, cute, 1girl, ahoge, ascot, blue eyes, blue gemstone, blue hair, blue headwear, blue jacket, gem, hair between eyes, hat, jacket, light blue hair, looking at viewer, multicolored hair, closed mouth, solo, top hat, white hair"
35
+ sample_negative = "Weighted"
lora_diffusers.py ADDED
@@ -0,0 +1,539 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ LoRA module for Diffusers
3
+ ==========================
4
+
5
+ This file works independently and is designed to operate with Diffusers.
6
+
7
+ Credits
8
+ -------
9
+ - Modified from: https://github.com/vladmandic/automatic/blob/master/modules/lora_diffusers.py
10
+ - Originally from: https://github.com/kohya-ss/sd-scripts/blob/sdxl/networks/lora_diffusers.py
11
+ """
12
+
13
+ import bisect
14
+ import math
15
+ import random
16
+ from typing import Any, Dict, List, Mapping, Optional, Union
17
+ from diffusers import UNet2DConditionModel
18
+ import numpy as np
19
+ from tqdm import tqdm
20
+ import diffusers.models.lora as diffusers_lora
21
+ from transformers import CLIPTextModel
22
+ import torch
23
+
24
+
25
+ def make_unet_conversion_map() -> Dict[str, str]:
26
+ unet_conversion_map_layer = []
27
+
28
+ for i in range(3): # num_blocks is 3 in sdxl
29
+ # loop over downblocks/upblocks
30
+ for j in range(2):
31
+ # loop over resnets/attentions for downblocks
32
+ hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
33
+ sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
34
+ unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
35
+
36
+ if i < 3:
37
+ # no attention layers in down_blocks.3
38
+ hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
39
+ sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."
40
+ unet_conversion_map_layer.append(
41
+ (sd_down_atn_prefix, hf_down_atn_prefix)
42
+ )
43
+
44
+ for j in range(3):
45
+ # loop over resnets/attentions for upblocks
46
+ hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
47
+ sd_up_res_prefix = f"output_blocks.{3*i + j}.0."
48
+ unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
49
+
50
+ # if i > 0: commentout for sdxl
51
+ # no attention layers in up_blocks.0
52
+ hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
53
+ sd_up_atn_prefix = f"output_blocks.{3*i + j}.1."
54
+ unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
55
+
56
+ if i < 3:
57
+ # no downsample in down_blocks.3
58
+ hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
59
+ sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."
60
+ unet_conversion_map_layer.append(
61
+ (sd_downsample_prefix, hf_downsample_prefix)
62
+ )
63
+
64
+ # no upsample in up_blocks.3
65
+ hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
66
+ sd_upsample_prefix = f"output_blocks.{3*i + 2}.{2}." # change for sdxl
67
+ unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
68
+
69
+ hf_mid_atn_prefix = "mid_block.attentions.0."
70
+ sd_mid_atn_prefix = "middle_block.1."
71
+ unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
72
+
73
+ for j in range(2):
74
+ hf_mid_res_prefix = f"mid_block.resnets.{j}."
75
+ sd_mid_res_prefix = f"middle_block.{2*j}."
76
+ unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
77
+
78
+ unet_conversion_map_resnet = [
79
+ # (stable-diffusion, HF Diffusers)
80
+ ("in_layers.0.", "norm1."),
81
+ ("in_layers.2.", "conv1."),
82
+ ("out_layers.0.", "norm2."),
83
+ ("out_layers.3.", "conv2."),
84
+ ("emb_layers.1.", "time_emb_proj."),
85
+ ("skip_connection.", "conv_shortcut."),
86
+ ]
87
+
88
+ unet_conversion_map = []
89
+ for sd, hf in unet_conversion_map_layer:
90
+ if "resnets" in hf:
91
+ for sd_res, hf_res in unet_conversion_map_resnet:
92
+ unet_conversion_map.append((sd + sd_res, hf + hf_res))
93
+ else:
94
+ unet_conversion_map.append((sd, hf))
95
+
96
+ for j in range(2):
97
+ hf_time_embed_prefix = f"time_embedding.linear_{j+1}."
98
+ sd_time_embed_prefix = f"time_embed.{j*2}."
99
+ unet_conversion_map.append((sd_time_embed_prefix, hf_time_embed_prefix))
100
+
101
+ for j in range(2):
102
+ hf_label_embed_prefix = f"add_embedding.linear_{j+1}."
103
+ sd_label_embed_prefix = f"label_emb.0.{j*2}."
104
+ unet_conversion_map.append((sd_label_embed_prefix, hf_label_embed_prefix))
105
+
106
+ unet_conversion_map.append(("input_blocks.0.0.", "conv_in."))
107
+ unet_conversion_map.append(("out.0.", "conv_norm_out."))
108
+ unet_conversion_map.append(("out.2.", "conv_out."))
109
+
110
+ sd_hf_conversion_map = {
111
+ sd.replace(".", "_")[:-1]: hf.replace(".", "_")[:-1]
112
+ for sd, hf in unet_conversion_map
113
+ }
114
+ return sd_hf_conversion_map
115
+
116
+
117
+ UNET_CONVERSION_MAP = make_unet_conversion_map()
118
+
119
+
120
+ class LoRAModule(torch.nn.Module):
121
+ """
122
+ replaces forward method of the original Linear, instead of replacing the original Linear module.
123
+ """
124
+
125
+ def __init__(
126
+ self,
127
+ lora_name,
128
+ org_module: torch.nn.Module,
129
+ multiplier=1.0,
130
+ lora_dim=4,
131
+ alpha=1,
132
+ ):
133
+ """if alpha == 0 or None, alpha is rank (no scaling)."""
134
+ super().__init__()
135
+ self.lora_name = lora_name
136
+
137
+ if isinstance(
138
+ org_module, diffusers_lora.LoRACompatibleConv
139
+ ): # Modified to support Diffusers>=0.19.2
140
+ in_dim = org_module.in_channels
141
+ out_dim = org_module.out_channels
142
+ else:
143
+ in_dim = org_module.in_features
144
+ out_dim = org_module.out_features
145
+
146
+ self.lora_dim = lora_dim
147
+
148
+ if isinstance(
149
+ org_module, diffusers_lora.LoRACompatibleConv
150
+ ): # Modified to support Diffusers>=0.19.2
151
+ kernel_size = org_module.kernel_size
152
+ stride = org_module.stride
153
+ padding = org_module.padding
154
+ self.lora_down = torch.nn.Conv2d(
155
+ in_dim, self.lora_dim, kernel_size, stride, padding, bias=False
156
+ )
157
+ self.lora_up = torch.nn.Conv2d(
158
+ self.lora_dim, out_dim, (1, 1), (1, 1), bias=False
159
+ )
160
+ else:
161
+ self.lora_down = torch.nn.Linear(in_dim, self.lora_dim, bias=False)
162
+ self.lora_up = torch.nn.Linear(self.lora_dim, out_dim, bias=False)
163
+
164
+ if isinstance(alpha, torch.Tensor):
165
+ alpha = alpha.detach().float().numpy() # without casting, bf16 causes error
166
+ alpha = self.lora_dim if alpha is None or alpha == 0 else alpha
167
+ self.scale = alpha / self.lora_dim
168
+ self.register_buffer(
169
+ "alpha", torch.tensor(alpha)
170
+ ) # 勾配計算に含めない / not included in gradient calculation
171
+
172
+ # same as microsoft's
173
+ torch.nn.init.kaiming_uniform_(self.lora_down.weight, a=math.sqrt(5))
174
+ torch.nn.init.zeros_(self.lora_up.weight)
175
+
176
+ self.multiplier = multiplier
177
+ self.org_module = [org_module]
178
+ self.enabled = True
179
+ self.network: LoRANetwork = None
180
+ self.org_forward = None
181
+
182
+ # override org_module's forward method
183
+ def apply_to(self, multiplier=None):
184
+ if multiplier is not None:
185
+ self.multiplier = multiplier
186
+ if self.org_forward is None:
187
+ self.org_forward = self.org_module[0].forward
188
+ self.org_module[0].forward = self.forward
189
+
190
+ # restore org_module's forward method
191
+ def unapply_to(self):
192
+ if self.org_forward is not None:
193
+ self.org_module[0].forward = self.org_forward
194
+
195
+ # forward with lora
196
+ def forward(self, x):
197
+ if not self.enabled:
198
+ return self.org_forward(x)
199
+ return (
200
+ self.org_forward(x)
201
+ + self.lora_up(self.lora_down(x)) * self.multiplier * self.scale
202
+ )
203
+
204
+ def set_network(self, network):
205
+ self.network = network
206
+
207
+ # merge lora weight to org weight
208
+ def merge_to(self, multiplier=1.0):
209
+ # get lora weight
210
+ lora_weight = self.get_weight(multiplier)
211
+
212
+ # get org weight
213
+ org_sd = self.org_module[0].state_dict()
214
+ org_weight = org_sd["weight"]
215
+ weight = org_weight + lora_weight.to(org_weight.device, dtype=org_weight.dtype)
216
+
217
+ # set weight to org_module
218
+ org_sd["weight"] = weight
219
+ self.org_module[0].load_state_dict(org_sd)
220
+
221
+ # restore org weight from lora weight
222
+ def restore_from(self, multiplier=1.0):
223
+ # get lora weight
224
+ lora_weight = self.get_weight(multiplier)
225
+
226
+ # get org weight
227
+ org_sd = self.org_module[0].state_dict()
228
+ org_weight = org_sd["weight"]
229
+ weight = org_weight - lora_weight.to(org_weight.device, dtype=org_weight.dtype)
230
+
231
+ # set weight to org_module
232
+ org_sd["weight"] = weight
233
+ self.org_module[0].load_state_dict(org_sd)
234
+
235
+ # return lora weight
236
+ def get_weight(self, multiplier=None):
237
+ if multiplier is None:
238
+ multiplier = self.multiplier
239
+
240
+ # get up/down weight from module
241
+ up_weight = self.lora_up.weight.to(torch.float)
242
+ down_weight = self.lora_down.weight.to(torch.float)
243
+
244
+ # pre-calculated weight
245
+ if len(down_weight.size()) == 2:
246
+ # linear
247
+ weight = self.multiplier * (up_weight @ down_weight) * self.scale
248
+ elif down_weight.size()[2:4] == (1, 1):
249
+ # conv2d 1x1
250
+ weight = (
251
+ self.multiplier
252
+ * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2))
253
+ .unsqueeze(2)
254
+ .unsqueeze(3)
255
+ * self.scale
256
+ )
257
+ else:
258
+ # conv2d 3x3
259
+ conved = torch.nn.functional.conv2d(
260
+ down_weight.permute(1, 0, 2, 3), up_weight
261
+ ).permute(1, 0, 2, 3)
262
+ weight = self.multiplier * conved * self.scale
263
+
264
+ return weight
265
+
266
+
267
+ # Create network from weights for inference, weights are not loaded here
268
+ def create_network_from_weights(
269
+ text_encoder: Union[CLIPTextModel, List[CLIPTextModel]],
270
+ unet: UNet2DConditionModel,
271
+ weights_sd: Dict,
272
+ multiplier: float = 1.0,
273
+ ):
274
+ # get dim/alpha mapping
275
+ modules_dim = {}
276
+ modules_alpha = {}
277
+ for key, value in weights_sd.items():
278
+ if "." not in key:
279
+ continue
280
+
281
+ lora_name = key.split(".")[0]
282
+ if "alpha" in key:
283
+ modules_alpha[lora_name] = value
284
+ elif "lora_down" in key:
285
+ dim = value.size()[0]
286
+ modules_dim[lora_name] = dim
287
+ # print(lora_name, value.size(), dim)
288
+
289
+ # support old LoRA without alpha
290
+ for key in modules_dim.keys():
291
+ if key not in modules_alpha:
292
+ modules_alpha[key] = modules_dim[key]
293
+
294
+ return LoRANetwork(
295
+ text_encoder,
296
+ unet,
297
+ multiplier=multiplier,
298
+ modules_dim=modules_dim,
299
+ modules_alpha=modules_alpha,
300
+ )
301
+
302
+
303
+ def merge_lora_weights(pipe, weights_sd: Dict, multiplier: float = 1.0):
304
+ text_encoders = (
305
+ [pipe.text_encoder, pipe.text_encoder_2]
306
+ if hasattr(pipe, "text_encoder_2")
307
+ else [pipe.text_encoder]
308
+ )
309
+ unet = pipe.unet
310
+
311
+ lora_network = create_network_from_weights(
312
+ text_encoders, unet, weights_sd, multiplier=multiplier
313
+ )
314
+ lora_network.load_state_dict(weights_sd)
315
+ lora_network.merge_to(multiplier=multiplier)
316
+
317
+
318
+ # block weightや学習に対応しない簡易版 / simple version without block weight and training
319
+ class LoRANetwork(torch.nn.Module):
320
+ UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel"]
321
+ UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = [
322
+ "ResnetBlock2D",
323
+ "Downsample2D",
324
+ "Upsample2D",
325
+ ]
326
+ TEXT_ENCODER_TARGET_REPLACE_MODULE = ["CLIPAttention", "CLIPMLP"]
327
+ LORA_PREFIX_UNET = "lora_unet"
328
+ LORA_PREFIX_TEXT_ENCODER = "lora_te"
329
+
330
+ # SDXL: must starts with LORA_PREFIX_TEXT_ENCODER
331
+ LORA_PREFIX_TEXT_ENCODER1 = "lora_te1"
332
+ LORA_PREFIX_TEXT_ENCODER2 = "lora_te2"
333
+
334
+ def __init__(
335
+ self,
336
+ text_encoder: Union[List[CLIPTextModel], CLIPTextModel],
337
+ unet: UNet2DConditionModel,
338
+ multiplier: float = 1.0,
339
+ modules_dim: Optional[Dict[str, int]] = None,
340
+ modules_alpha: Optional[Dict[str, int]] = None,
341
+ varbose: Optional[bool] = False,
342
+ ) -> None:
343
+ super().__init__()
344
+ self.multiplier = multiplier
345
+
346
+ print(f"create LoRA network from weights")
347
+
348
+ # convert SDXL Stability AI's U-Net modules to Diffusers
349
+ converted = self.convert_unet_modules(modules_dim, modules_alpha)
350
+ if converted:
351
+ print(
352
+ f"converted {converted} Stability AI's U-Net LoRA modules to Diffusers (SDXL)"
353
+ )
354
+
355
+ # create module instances
356
+ def create_modules(
357
+ is_unet: bool,
358
+ text_encoder_idx: Optional[int], # None, 1, 2
359
+ root_module: torch.nn.Module,
360
+ target_replace_modules: List[torch.nn.Module],
361
+ ) -> List[LoRAModule]:
362
+ prefix = (
363
+ self.LORA_PREFIX_UNET
364
+ if is_unet
365
+ else (
366
+ self.LORA_PREFIX_TEXT_ENCODER
367
+ if text_encoder_idx is None
368
+ else (
369
+ self.LORA_PREFIX_TEXT_ENCODER1
370
+ if text_encoder_idx == 1
371
+ else self.LORA_PREFIX_TEXT_ENCODER2
372
+ )
373
+ )
374
+ )
375
+ loras = []
376
+ skipped = []
377
+ for name, module in root_module.named_modules():
378
+ if module.__class__.__name__ in target_replace_modules:
379
+ for child_name, child_module in module.named_modules():
380
+ is_linear = isinstance(
381
+ child_module,
382
+ (torch.nn.Linear, diffusers_lora.LoRACompatibleLinear),
383
+ ) # Modified to support Diffusers>=0.19.2
384
+ is_conv2d = isinstance(
385
+ child_module,
386
+ (torch.nn.Conv2d, diffusers_lora.LoRACompatibleConv),
387
+ ) # Modified to support Diffusers>=0.19.2
388
+
389
+ if is_linear or is_conv2d:
390
+ lora_name = prefix + "." + name + "." + child_name
391
+ lora_name = lora_name.replace(".", "_")
392
+
393
+ if lora_name not in modules_dim:
394
+ # print(f"skipped {lora_name} (not found in modules_dim)")
395
+ skipped.append(lora_name)
396
+ continue
397
+
398
+ dim = modules_dim[lora_name]
399
+ alpha = modules_alpha[lora_name]
400
+ lora = LoRAModule(
401
+ lora_name,
402
+ child_module,
403
+ self.multiplier,
404
+ dim,
405
+ alpha,
406
+ )
407
+ loras.append(lora)
408
+ return loras, skipped
409
+
410
+ text_encoders = text_encoder if type(text_encoder) == list else [text_encoder]
411
+
412
+ # create LoRA for text encoder
413
+ # 毎回すべてのモジュールを作るのは無駄なので要検討 / it is wasteful to create all modules every time, need to consider
414
+ self.text_encoder_loras: List[LoRAModule] = []
415
+ skipped_te = []
416
+ for i, text_encoder in enumerate(text_encoders):
417
+ if len(text_encoders) > 1:
418
+ index = i + 1
419
+ else:
420
+ index = None
421
+
422
+ text_encoder_loras, skipped = create_modules(
423
+ False,
424
+ index,
425
+ text_encoder,
426
+ LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE,
427
+ )
428
+ self.text_encoder_loras.extend(text_encoder_loras)
429
+ skipped_te += skipped
430
+ print(f"create LoRA for Text Encoder: {len(self.text_encoder_loras)} modules.")
431
+ if len(skipped_te) > 0:
432
+ print(f"skipped {len(skipped_te)} modules because of missing weight.")
433
+
434
+ # extend U-Net target modules to include Conv2d 3x3
435
+ target_modules = (
436
+ LoRANetwork.UNET_TARGET_REPLACE_MODULE
437
+ + LoRANetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3
438
+ )
439
+
440
+ self.unet_loras: List[LoRAModule]
441
+ self.unet_loras, skipped_un = create_modules(True, None, unet, target_modules)
442
+ print(f"create LoRA for U-Net: {len(self.unet_loras)} modules.")
443
+ if len(skipped_un) > 0:
444
+ print(f"skipped {len(skipped_un)} modules because of missing weight.")
445
+
446
+ # assertion
447
+ names = set()
448
+ for lora in self.text_encoder_loras + self.unet_loras:
449
+ names.add(lora.lora_name)
450
+ for lora_name in modules_dim.keys():
451
+ assert (
452
+ lora_name in names
453
+ ), f"{lora_name} is not found in created LoRA modules."
454
+
455
+ # make to work load_state_dict
456
+ for lora in self.text_encoder_loras + self.unet_loras:
457
+ self.add_module(lora.lora_name, lora)
458
+
459
+ # SDXL: convert SDXL Stability AI's U-Net modules to Diffusers
460
+ def convert_unet_modules(self, modules_dim, modules_alpha):
461
+ converted_count = 0
462
+ not_converted_count = 0
463
+
464
+ map_keys = list(UNET_CONVERSION_MAP.keys())
465
+ map_keys.sort()
466
+
467
+ for key in list(modules_dim.keys()):
468
+ if key.startswith(LoRANetwork.LORA_PREFIX_UNET + "_"):
469
+ search_key = key.replace(LoRANetwork.LORA_PREFIX_UNET + "_", "")
470
+ position = bisect.bisect_right(map_keys, search_key)
471
+ map_key = map_keys[position - 1]
472
+ if search_key.startswith(map_key):
473
+ new_key = key.replace(map_key, UNET_CONVERSION_MAP[map_key])
474
+ modules_dim[new_key] = modules_dim[key]
475
+ modules_alpha[new_key] = modules_alpha[key]
476
+ del modules_dim[key]
477
+ del modules_alpha[key]
478
+ converted_count += 1
479
+ else:
480
+ not_converted_count += 1
481
+ assert (
482
+ converted_count == 0 or not_converted_count == 0
483
+ ), f"some modules are not converted: {converted_count} converted, {not_converted_count} not converted"
484
+ return converted_count
485
+
486
+ def set_multiplier(self, multiplier):
487
+ self.multiplier = multiplier
488
+ for lora in self.text_encoder_loras + self.unet_loras:
489
+ lora.multiplier = self.multiplier
490
+
491
+ def apply_to(self, multiplier=1.0, apply_text_encoder=True, apply_unet=True):
492
+ if apply_text_encoder:
493
+ print("enable LoRA for text encoder")
494
+ for lora in self.text_encoder_loras:
495
+ lora.apply_to(multiplier)
496
+ if apply_unet:
497
+ print("enable LoRA for U-Net")
498
+ for lora in self.unet_loras:
499
+ lora.apply_to(multiplier)
500
+
501
+ def unapply_to(self):
502
+ for lora in self.text_encoder_loras + self.unet_loras:
503
+ lora.unapply_to()
504
+
505
+ def merge_to(self, multiplier=1.0):
506
+ print("merge LoRA weights to original weights")
507
+ for lora in tqdm(self.text_encoder_loras + self.unet_loras):
508
+ lora.merge_to(multiplier)
509
+ print(f"weights are merged")
510
+
511
+ def restore_from(self, multiplier=1.0):
512
+ print("restore LoRA weights from original weights")
513
+ for lora in tqdm(self.text_encoder_loras + self.unet_loras):
514
+ lora.restore_from(multiplier)
515
+ print(f"weights are restored")
516
+
517
+ def load_state_dict(self, state_dict: Mapping[str, Any], strict: bool = True):
518
+ # convert SDXL Stability AI's state dict to Diffusers' based state dict
519
+ map_keys = list(UNET_CONVERSION_MAP.keys()) # prefix of U-Net modules
520
+ map_keys.sort()
521
+ for key in list(state_dict.keys()):
522
+ if key.startswith(LoRANetwork.LORA_PREFIX_UNET + "_"):
523
+ search_key = key.replace(LoRANetwork.LORA_PREFIX_UNET + "_", "")
524
+ position = bisect.bisect_right(map_keys, search_key)
525
+ map_key = map_keys[position - 1]
526
+ if search_key.startswith(map_key):
527
+ new_key = key.replace(map_key, UNET_CONVERSION_MAP[map_key])
528
+ state_dict[new_key] = state_dict[key]
529
+ del state_dict[key]
530
+
531
+ # in case of V2, some weights have different shape, so we need to convert them
532
+ # because V2 LoRA is based on U-Net created by use_linear_projection=False
533
+ my_state_dict = self.state_dict()
534
+ for key in state_dict.keys():
535
+ if state_dict[key].size() != my_state_dict[key].size():
536
+ # print(f"convert {key} from {state_dict[key].size()} to {my_state_dict[key].size()}")
537
+ state_dict[key] = state_dict[key].view(my_state_dict[key].size())
538
+
539
+ return super().load_state_dict(state_dict, strict)
lpw_stable_diffusion_xl.py ADDED
@@ -0,0 +1,1496 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ## ----------------------------------------------------------
2
+ # A SDXL pipeline can take unlimited weighted prompt
3
+ #
4
+ # Author: Andrew Zhu
5
+ # Github: https://github.com/xhinker
6
+ # Medium: https://medium.com/@xhinker
7
+ ## -----------------------------------------------------------
8
+
9
+ import inspect
10
+ import os
11
+ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
12
+
13
+ import torch
14
+ from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
15
+
16
+ from diffusers import DiffusionPipeline, StableDiffusionXLPipeline
17
+ from diffusers.image_processor import VaeImageProcessor
18
+ from diffusers.loaders import (
19
+ FromSingleFileMixin,
20
+ LoraLoaderMixin,
21
+ TextualInversionLoaderMixin,
22
+ )
23
+ from diffusers.models import AutoencoderKL, UNet2DConditionModel
24
+ from diffusers.models.attention_processor import (
25
+ AttnProcessor2_0,
26
+ LoRAAttnProcessor2_0,
27
+ LoRAXFormersAttnProcessor,
28
+ XFormersAttnProcessor,
29
+ )
30
+ from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
31
+ from diffusers.schedulers import KarrasDiffusionSchedulers
32
+ from diffusers.utils import (
33
+ is_accelerate_available,
34
+ is_accelerate_version,
35
+ is_invisible_watermark_available,
36
+ logging,
37
+ randn_tensor,
38
+ replace_example_docstring,
39
+ )
40
+
41
+
42
+ if is_invisible_watermark_available():
43
+ from diffusers.pipelines.stable_diffusion_xl.watermark import (
44
+ StableDiffusionXLWatermarker,
45
+ )
46
+
47
+
48
+ def parse_prompt_attention(text):
49
+ """
50
+ Parses a string with attention tokens and returns a list of pairs: text and its associated weight.
51
+ Accepted tokens are:
52
+ (abc) - increases attention to abc by a multiplier of 1.1
53
+ (abc:3.12) - increases attention to abc by a multiplier of 3.12
54
+ [abc] - decreases attention to abc by a multiplier of 1.1
55
+ \( - literal character '('
56
+ \[ - literal character '['
57
+ \) - literal character ')'
58
+ \] - literal character ']'
59
+ \\ - literal character '\'
60
+ anything else - just text
61
+
62
+ >>> parse_prompt_attention('normal text')
63
+ [['normal text', 1.0]]
64
+ >>> parse_prompt_attention('an (important) word')
65
+ [['an ', 1.0], ['important', 1.1], [' word', 1.0]]
66
+ >>> parse_prompt_attention('(unbalanced')
67
+ [['unbalanced', 1.1]]
68
+ >>> parse_prompt_attention('\(literal\]')
69
+ [['(literal]', 1.0]]
70
+ >>> parse_prompt_attention('(unnecessary)(parens)')
71
+ [['unnecessaryparens', 1.1]]
72
+ >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).')
73
+ [['a ', 1.0],
74
+ ['house', 1.5730000000000004],
75
+ [' ', 1.1],
76
+ ['on', 1.0],
77
+ [' a ', 1.1],
78
+ ['hill', 0.55],
79
+ [', sun, ', 1.1],
80
+ ['sky', 1.4641000000000006],
81
+ ['.', 1.1]]
82
+ """
83
+ import re
84
+
85
+ re_attention = re.compile(
86
+ r"""
87
+ \\\(|\\\)|\\\[|\\]|\\\\|\\|\(|\[|:([+-]?[.\d]+)\)|
88
+ \)|]|[^\\()\[\]:]+|:
89
+ """,
90
+ re.X,
91
+ )
92
+
93
+ re_break = re.compile(r"\s*\bBREAK\b\s*", re.S)
94
+
95
+ res = []
96
+ round_brackets = []
97
+ square_brackets = []
98
+
99
+ round_bracket_multiplier = 1.1
100
+ square_bracket_multiplier = 1 / 1.1
101
+
102
+ def multiply_range(start_position, multiplier):
103
+ for p in range(start_position, len(res)):
104
+ res[p][1] *= multiplier
105
+
106
+ for m in re_attention.finditer(text):
107
+ text = m.group(0)
108
+ weight = m.group(1)
109
+
110
+ if text.startswith("\\"):
111
+ res.append([text[1:], 1.0])
112
+ elif text == "(":
113
+ round_brackets.append(len(res))
114
+ elif text == "[":
115
+ square_brackets.append(len(res))
116
+ elif weight is not None and len(round_brackets) > 0:
117
+ multiply_range(round_brackets.pop(), float(weight))
118
+ elif text == ")" and len(round_brackets) > 0:
119
+ multiply_range(round_brackets.pop(), round_bracket_multiplier)
120
+ elif text == "]" and len(square_brackets) > 0:
121
+ multiply_range(square_brackets.pop(), square_bracket_multiplier)
122
+ else:
123
+ parts = re.split(re_break, text)
124
+ for i, part in enumerate(parts):
125
+ if i > 0:
126
+ res.append(["BREAK", -1])
127
+ res.append([part, 1.0])
128
+
129
+ for pos in round_brackets:
130
+ multiply_range(pos, round_bracket_multiplier)
131
+
132
+ for pos in square_brackets:
133
+ multiply_range(pos, square_bracket_multiplier)
134
+
135
+ if len(res) == 0:
136
+ res = [["", 1.0]]
137
+
138
+ # merge runs of identical weights
139
+ i = 0
140
+ while i + 1 < len(res):
141
+ if res[i][1] == res[i + 1][1]:
142
+ res[i][0] += res[i + 1][0]
143
+ res.pop(i + 1)
144
+ else:
145
+ i += 1
146
+
147
+ return res
148
+
149
+
150
+ def get_prompts_tokens_with_weights(clip_tokenizer: CLIPTokenizer, prompt: str):
151
+ """
152
+ Get prompt token ids and weights, this function works for both prompt and negative prompt
153
+
154
+ Args:
155
+ pipe (CLIPTokenizer)
156
+ A CLIPTokenizer
157
+ prompt (str)
158
+ A prompt string with weights
159
+
160
+ Returns:
161
+ text_tokens (list)
162
+ A list contains token ids
163
+ text_weight (list)
164
+ A list contains the correspodent weight of token ids
165
+
166
+ Example:
167
+ import torch
168
+ from transformers import CLIPTokenizer
169
+
170
+ clip_tokenizer = CLIPTokenizer.from_pretrained(
171
+ "stablediffusionapi/deliberate-v2"
172
+ , subfolder = "tokenizer"
173
+ , dtype = torch.float16
174
+ )
175
+
176
+ token_id_list, token_weight_list = get_prompts_tokens_with_weights(
177
+ clip_tokenizer = clip_tokenizer
178
+ ,prompt = "a (red:1.5) cat"*70
179
+ )
180
+ """
181
+ texts_and_weights = parse_prompt_attention(prompt)
182
+ text_tokens, text_weights = [], []
183
+ for word, weight in texts_and_weights:
184
+ # tokenize and discard the starting and the ending token
185
+ token = clip_tokenizer(word, truncation=False).input_ids[
186
+ 1:-1
187
+ ] # so that tokenize whatever length prompt
188
+ # the returned token is a 1d list: [320, 1125, 539, 320]
189
+
190
+ # merge the new tokens to the all tokens holder: text_tokens
191
+ text_tokens = [*text_tokens, *token]
192
+
193
+ # each token chunk will come with one weight, like ['red cat', 2.0]
194
+ # need to expand weight for each token.
195
+ chunk_weights = [weight] * len(token)
196
+
197
+ # append the weight back to the weight holder: text_weights
198
+ text_weights = [*text_weights, *chunk_weights]
199
+ return text_tokens, text_weights
200
+
201
+
202
+ def group_tokens_and_weights(token_ids: list, weights: list, pad_last_block=False):
203
+ """
204
+ Produce tokens and weights in groups and pad the missing tokens
205
+
206
+ Args:
207
+ token_ids (list)
208
+ The token ids from tokenizer
209
+ weights (list)
210
+ The weights list from function get_prompts_tokens_with_weights
211
+ pad_last_block (bool)
212
+ Control if fill the last token list to 75 tokens with eos
213
+ Returns:
214
+ new_token_ids (2d list)
215
+ new_weights (2d list)
216
+
217
+ Example:
218
+ token_groups,weight_groups = group_tokens_and_weights(
219
+ token_ids = token_id_list
220
+ , weights = token_weight_list
221
+ )
222
+ """
223
+ bos, eos = 49406, 49407
224
+
225
+ # this will be a 2d list
226
+ new_token_ids = []
227
+ new_weights = []
228
+ while len(token_ids) >= 75:
229
+ # get the first 75 tokens
230
+ head_75_tokens = [token_ids.pop(0) for _ in range(75)]
231
+ head_75_weights = [weights.pop(0) for _ in range(75)]
232
+
233
+ # extract token ids and weights
234
+ temp_77_token_ids = [bos] + head_75_tokens + [eos]
235
+ temp_77_weights = [1.0] + head_75_weights + [1.0]
236
+
237
+ # add 77 token and weights chunk to the holder list
238
+ new_token_ids.append(temp_77_token_ids)
239
+ new_weights.append(temp_77_weights)
240
+
241
+ # padding the left
242
+ if len(token_ids) > 0:
243
+ padding_len = 75 - len(token_ids) if pad_last_block else 0
244
+
245
+ temp_77_token_ids = [bos] + token_ids + [eos] * padding_len + [eos]
246
+ new_token_ids.append(temp_77_token_ids)
247
+
248
+ temp_77_weights = [1.0] + weights + [1.0] * padding_len + [1.0]
249
+ new_weights.append(temp_77_weights)
250
+
251
+ return new_token_ids, new_weights
252
+
253
+
254
+ def get_weighted_text_embeddings_sdxl(
255
+ pipe: StableDiffusionXLPipeline,
256
+ prompt: str = "",
257
+ prompt_2: str = None,
258
+ neg_prompt: str = "",
259
+ neg_prompt_2: str = None,
260
+ ):
261
+ """
262
+ This function can process long prompt with weights, no length limitation
263
+ for Stable Diffusion XL
264
+
265
+ Args:
266
+ pipe (StableDiffusionPipeline)
267
+ prompt (str)
268
+ prompt_2 (str)
269
+ neg_prompt (str)
270
+ neg_prompt_2 (str)
271
+ Returns:
272
+ prompt_embeds (torch.Tensor)
273
+ neg_prompt_embeds (torch.Tensor)
274
+ """
275
+ if prompt_2:
276
+ prompt = f"{prompt} {prompt_2}"
277
+
278
+ if neg_prompt_2:
279
+ neg_prompt = f"{neg_prompt} {neg_prompt_2}"
280
+
281
+ eos = pipe.tokenizer.eos_token_id
282
+
283
+ # tokenizer 1
284
+ prompt_tokens, prompt_weights = get_prompts_tokens_with_weights(
285
+ pipe.tokenizer, prompt
286
+ )
287
+
288
+ neg_prompt_tokens, neg_prompt_weights = get_prompts_tokens_with_weights(
289
+ pipe.tokenizer, neg_prompt
290
+ )
291
+
292
+ # tokenizer 2
293
+ prompt_tokens_2, prompt_weights_2 = get_prompts_tokens_with_weights(
294
+ pipe.tokenizer_2, prompt
295
+ )
296
+
297
+ neg_prompt_tokens_2, neg_prompt_weights_2 = get_prompts_tokens_with_weights(
298
+ pipe.tokenizer_2, neg_prompt
299
+ )
300
+
301
+ # padding the shorter one for prompt set 1
302
+ prompt_token_len = len(prompt_tokens)
303
+ neg_prompt_token_len = len(neg_prompt_tokens)
304
+
305
+ if prompt_token_len > neg_prompt_token_len:
306
+ # padding the neg_prompt with eos token
307
+ neg_prompt_tokens = neg_prompt_tokens + [eos] * abs(
308
+ prompt_token_len - neg_prompt_token_len
309
+ )
310
+ neg_prompt_weights = neg_prompt_weights + [1.0] * abs(
311
+ prompt_token_len - neg_prompt_token_len
312
+ )
313
+ else:
314
+ # padding the prompt
315
+ prompt_tokens = prompt_tokens + [eos] * abs(
316
+ prompt_token_len - neg_prompt_token_len
317
+ )
318
+ prompt_weights = prompt_weights + [1.0] * abs(
319
+ prompt_token_len - neg_prompt_token_len
320
+ )
321
+
322
+ # padding the shorter one for token set 2
323
+ prompt_token_len_2 = len(prompt_tokens_2)
324
+ neg_prompt_token_len_2 = len(neg_prompt_tokens_2)
325
+
326
+ if prompt_token_len_2 > neg_prompt_token_len_2:
327
+ # padding the neg_prompt with eos token
328
+ neg_prompt_tokens_2 = neg_prompt_tokens_2 + [eos] * abs(
329
+ prompt_token_len_2 - neg_prompt_token_len_2
330
+ )
331
+ neg_prompt_weights_2 = neg_prompt_weights_2 + [1.0] * abs(
332
+ prompt_token_len_2 - neg_prompt_token_len_2
333
+ )
334
+ else:
335
+ # padding the prompt
336
+ prompt_tokens_2 = prompt_tokens_2 + [eos] * abs(
337
+ prompt_token_len_2 - neg_prompt_token_len_2
338
+ )
339
+ prompt_weights_2 = prompt_weights + [1.0] * abs(
340
+ prompt_token_len_2 - neg_prompt_token_len_2
341
+ )
342
+
343
+ embeds = []
344
+ neg_embeds = []
345
+
346
+ prompt_token_groups, prompt_weight_groups = group_tokens_and_weights(
347
+ prompt_tokens.copy(), prompt_weights.copy()
348
+ )
349
+
350
+ neg_prompt_token_groups, neg_prompt_weight_groups = group_tokens_and_weights(
351
+ neg_prompt_tokens.copy(), neg_prompt_weights.copy()
352
+ )
353
+
354
+ prompt_token_groups_2, prompt_weight_groups_2 = group_tokens_and_weights(
355
+ prompt_tokens_2.copy(), prompt_weights_2.copy()
356
+ )
357
+
358
+ neg_prompt_token_groups_2, neg_prompt_weight_groups_2 = group_tokens_and_weights(
359
+ neg_prompt_tokens_2.copy(), neg_prompt_weights_2.copy()
360
+ )
361
+
362
+ # get prompt embeddings one by one is not working.
363
+ for i in range(len(prompt_token_groups)):
364
+ # get positive prompt embeddings with weights
365
+ token_tensor = torch.tensor(
366
+ [prompt_token_groups[i]], dtype=torch.long, device=pipe.device
367
+ )
368
+ weight_tensor = torch.tensor(
369
+ prompt_weight_groups[i], dtype=torch.float16, device=pipe.device
370
+ )
371
+
372
+ token_tensor_2 = torch.tensor(
373
+ [prompt_token_groups_2[i]], dtype=torch.long, device=pipe.device
374
+ )
375
+
376
+ # use first text encoder
377
+ prompt_embeds_1 = pipe.text_encoder(
378
+ token_tensor.to(pipe.device), output_hidden_states=True
379
+ )
380
+ prompt_embeds_1_hidden_states = prompt_embeds_1.hidden_states[-2]
381
+
382
+ # use second text encoder
383
+ prompt_embeds_2 = pipe.text_encoder_2(
384
+ token_tensor_2.to(pipe.device), output_hidden_states=True
385
+ )
386
+ prompt_embeds_2_hidden_states = prompt_embeds_2.hidden_states[-2]
387
+ pooled_prompt_embeds = prompt_embeds_2[0]
388
+
389
+ prompt_embeds_list = [
390
+ prompt_embeds_1_hidden_states,
391
+ prompt_embeds_2_hidden_states,
392
+ ]
393
+ token_embedding = torch.concat(prompt_embeds_list, dim=-1).squeeze(0)
394
+
395
+ for j in range(len(weight_tensor)):
396
+ if weight_tensor[j] != 1.0:
397
+ token_embedding[j] = (
398
+ token_embedding[-1]
399
+ + (token_embedding[j] - token_embedding[-1]) * weight_tensor[j]
400
+ )
401
+
402
+ token_embedding = token_embedding.unsqueeze(0)
403
+ embeds.append(token_embedding)
404
+
405
+ # get negative prompt embeddings with weights
406
+ neg_token_tensor = torch.tensor(
407
+ [neg_prompt_token_groups[i]], dtype=torch.long, device=pipe.device
408
+ )
409
+ neg_token_tensor_2 = torch.tensor(
410
+ [neg_prompt_token_groups_2[i]], dtype=torch.long, device=pipe.device
411
+ )
412
+ neg_weight_tensor = torch.tensor(
413
+ neg_prompt_weight_groups[i], dtype=torch.float16, device=pipe.device
414
+ )
415
+
416
+ # use first text encoder
417
+ neg_prompt_embeds_1 = pipe.text_encoder(
418
+ neg_token_tensor.to(pipe.device), output_hidden_states=True
419
+ )
420
+ neg_prompt_embeds_1_hidden_states = neg_prompt_embeds_1.hidden_states[-2]
421
+
422
+ # use second text encoder
423
+ neg_prompt_embeds_2 = pipe.text_encoder_2(
424
+ neg_token_tensor_2.to(pipe.device), output_hidden_states=True
425
+ )
426
+ neg_prompt_embeds_2_hidden_states = neg_prompt_embeds_2.hidden_states[-2]
427
+ negative_pooled_prompt_embeds = neg_prompt_embeds_2[0]
428
+
429
+ neg_prompt_embeds_list = [
430
+ neg_prompt_embeds_1_hidden_states,
431
+ neg_prompt_embeds_2_hidden_states,
432
+ ]
433
+ neg_token_embedding = torch.concat(neg_prompt_embeds_list, dim=-1).squeeze(0)
434
+
435
+ for z in range(len(neg_weight_tensor)):
436
+ if neg_weight_tensor[z] != 1.0:
437
+ neg_token_embedding[z] = (
438
+ neg_token_embedding[-1]
439
+ + (neg_token_embedding[z] - neg_token_embedding[-1])
440
+ * neg_weight_tensor[z]
441
+ )
442
+
443
+ neg_token_embedding = neg_token_embedding.unsqueeze(0)
444
+ neg_embeds.append(neg_token_embedding)
445
+
446
+ prompt_embeds = torch.cat(embeds, dim=1)
447
+ negative_prompt_embeds = torch.cat(neg_embeds, dim=1)
448
+
449
+ return (
450
+ prompt_embeds,
451
+ negative_prompt_embeds,
452
+ pooled_prompt_embeds,
453
+ negative_pooled_prompt_embeds,
454
+ )
455
+
456
+
457
+ # -------------------------------------------------------------------------------------------------------------------------------
458
+ # reuse the backbone code from StableDiffusionXLPipeline
459
+ # -------------------------------------------------------------------------------------------------------------------------------
460
+
461
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
462
+
463
+ EXAMPLE_DOC_STRING = """
464
+ Examples:
465
+ ```py
466
+ from diffusers import DiffusionPipeline
467
+ import torch
468
+
469
+ pipe = DiffusionPipeline.from_pretrained(
470
+ "stabilityai/stable-diffusion-xl-base-1.0"
471
+ , torch_dtype = torch.float16
472
+ , use_safetensors = True
473
+ , variant = "fp16"
474
+ , custom_pipeline = "lpw_stable_diffusion_xl",
475
+ )
476
+
477
+ prompt = "a white cat running on the grass"*20
478
+ prompt2 = "play a football"*20
479
+ prompt = f"{prompt},{prompt2}"
480
+ neg_prompt = "blur, low quality"
481
+
482
+ pipe.to("cuda")
483
+ images = pipe(
484
+ prompt = prompt
485
+ , negative_prompt = neg_prompt
486
+ ).images[0]
487
+
488
+ pipe.to("cpu")
489
+ torch.cuda.empty_cache()
490
+ images
491
+ ```
492
+ """
493
+
494
+
495
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
496
+ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
497
+ """
498
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
499
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
500
+ """
501
+ std_text = noise_pred_text.std(
502
+ dim=list(range(1, noise_pred_text.ndim)), keepdim=True
503
+ )
504
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
505
+ # rescale the results from guidance (fixes overexposure)
506
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
507
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
508
+ noise_cfg = (
509
+ guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
510
+ )
511
+ return noise_cfg
512
+
513
+
514
+ class SDXLLongPromptWeightingPipeline(
515
+ DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin
516
+ ):
517
+ r"""
518
+ Pipeline for text-to-image generation using Stable Diffusion XL.
519
+
520
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
521
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
522
+
523
+ In addition the pipeline inherits the following loading methods:
524
+ - *LoRA*: [`StableDiffusionXLPipeline.load_lora_weights`]
525
+ - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
526
+
527
+ as well as the following saving methods:
528
+ - *LoRA*: [`loaders.StableDiffusionXLPipeline.save_lora_weights`]
529
+
530
+ Args:
531
+ vae ([`AutoencoderKL`]):
532
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
533
+ text_encoder ([`CLIPTextModel`]):
534
+ Frozen text-encoder. Stable Diffusion XL uses the text portion of
535
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
536
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
537
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
538
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
539
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
540
+ specifically the
541
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
542
+ variant.
543
+ tokenizer (`CLIPTokenizer`):
544
+ Tokenizer of class
545
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
546
+ tokenizer_2 (`CLIPTokenizer`):
547
+ Second Tokenizer of class
548
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
549
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
550
+ scheduler ([`SchedulerMixin`]):
551
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
552
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
553
+ """
554
+
555
+ def __init__(
556
+ self,
557
+ vae: AutoencoderKL,
558
+ text_encoder: CLIPTextModel,
559
+ text_encoder_2: CLIPTextModelWithProjection,
560
+ tokenizer: CLIPTokenizer,
561
+ tokenizer_2: CLIPTokenizer,
562
+ unet: UNet2DConditionModel,
563
+ scheduler: KarrasDiffusionSchedulers,
564
+ force_zeros_for_empty_prompt: bool = True,
565
+ add_watermarker: Optional[bool] = None,
566
+ ):
567
+ super().__init__()
568
+
569
+ self.register_modules(
570
+ vae=vae,
571
+ text_encoder=text_encoder,
572
+ text_encoder_2=text_encoder_2,
573
+ tokenizer=tokenizer,
574
+ tokenizer_2=tokenizer_2,
575
+ unet=unet,
576
+ scheduler=scheduler,
577
+ )
578
+ self.register_to_config(
579
+ force_zeros_for_empty_prompt=force_zeros_for_empty_prompt
580
+ )
581
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
582
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
583
+ self.default_sample_size = self.unet.config.sample_size
584
+
585
+ add_watermarker = (
586
+ add_watermarker
587
+ if add_watermarker is not None
588
+ else is_invisible_watermark_available()
589
+ )
590
+
591
+ if add_watermarker:
592
+ self.watermark = StableDiffusionXLWatermarker()
593
+ else:
594
+ self.watermark = None
595
+
596
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
597
+ def enable_vae_slicing(self):
598
+ r"""
599
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
600
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
601
+ """
602
+ self.vae.enable_slicing()
603
+
604
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
605
+ def disable_vae_slicing(self):
606
+ r"""
607
+ Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
608
+ computing decoding in one step.
609
+ """
610
+ self.vae.disable_slicing()
611
+
612
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
613
+ def enable_vae_tiling(self):
614
+ r"""
615
+ Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
616
+ compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
617
+ processing larger images.
618
+ """
619
+ self.vae.enable_tiling()
620
+
621
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
622
+ def disable_vae_tiling(self):
623
+ r"""
624
+ Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
625
+ computing decoding in one step.
626
+ """
627
+ self.vae.disable_tiling()
628
+
629
+ def enable_model_cpu_offload(self, gpu_id=0):
630
+ r"""
631
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
632
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
633
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
634
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
635
+ """
636
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
637
+ from accelerate import cpu_offload_with_hook
638
+ else:
639
+ raise ImportError(
640
+ "`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher."
641
+ )
642
+
643
+ device = torch.device(f"cuda:{gpu_id}")
644
+
645
+ if self.device.type != "cpu":
646
+ self.to("cpu", silence_dtype_warnings=True)
647
+ torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
648
+
649
+ model_sequence = (
650
+ [self.text_encoder, self.text_encoder_2]
651
+ if self.text_encoder is not None
652
+ else [self.text_encoder_2]
653
+ )
654
+ model_sequence.extend([self.unet, self.vae])
655
+
656
+ hook = None
657
+ for cpu_offloaded_model in model_sequence:
658
+ _, hook = cpu_offload_with_hook(
659
+ cpu_offloaded_model, device, prev_module_hook=hook
660
+ )
661
+
662
+ # We'll offload the last model manually.
663
+ self.final_offload_hook = hook
664
+
665
+ def encode_prompt(
666
+ self,
667
+ prompt: str,
668
+ prompt_2: Optional[str] = None,
669
+ device: Optional[torch.device] = None,
670
+ num_images_per_prompt: int = 1,
671
+ do_classifier_free_guidance: bool = True,
672
+ negative_prompt: Optional[str] = None,
673
+ negative_prompt_2: Optional[str] = None,
674
+ prompt_embeds: Optional[torch.FloatTensor] = None,
675
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
676
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
677
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
678
+ lora_scale: Optional[float] = None,
679
+ ):
680
+ r"""
681
+ Encodes the prompt into text encoder hidden states.
682
+
683
+ Args:
684
+ prompt (`str` or `List[str]`, *optional*):
685
+ prompt to be encoded
686
+ prompt_2 (`str` or `List[str]`, *optional*):
687
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
688
+ used in both text-encoders
689
+ device: (`torch.device`):
690
+ torch device
691
+ num_images_per_prompt (`int`):
692
+ number of images that should be generated per prompt
693
+ do_classifier_free_guidance (`bool`):
694
+ whether to use classifier free guidance or not
695
+ negative_prompt (`str` or `List[str]`, *optional*):
696
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
697
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
698
+ less than `1`).
699
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
700
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
701
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
702
+ prompt_embeds (`torch.FloatTensor`, *optional*):
703
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
704
+ provided, text embeddings will be generated from `prompt` input argument.
705
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
706
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
707
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
708
+ argument.
709
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
710
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
711
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
712
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
713
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
714
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
715
+ input argument.
716
+ lora_scale (`float`, *optional*):
717
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
718
+ """
719
+ device = device or self._execution_device
720
+
721
+ # set lora scale so that monkey patched LoRA
722
+ # function of text encoder can correctly access it
723
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
724
+ self._lora_scale = lora_scale
725
+
726
+ if prompt is not None and isinstance(prompt, str):
727
+ batch_size = 1
728
+ elif prompt is not None and isinstance(prompt, list):
729
+ batch_size = len(prompt)
730
+ else:
731
+ batch_size = prompt_embeds.shape[0]
732
+
733
+ # Define tokenizers and text encoders
734
+ tokenizers = (
735
+ [self.tokenizer, self.tokenizer_2]
736
+ if self.tokenizer is not None
737
+ else [self.tokenizer_2]
738
+ )
739
+ text_encoders = (
740
+ [self.text_encoder, self.text_encoder_2]
741
+ if self.text_encoder is not None
742
+ else [self.text_encoder_2]
743
+ )
744
+
745
+ if prompt_embeds is None:
746
+ prompt_2 = prompt_2 or prompt
747
+ # textual inversion: procecss multi-vector tokens if necessary
748
+ prompt_embeds_list = []
749
+ prompts = [prompt, prompt_2]
750
+ for prompt, tokenizer, text_encoder in zip(
751
+ prompts, tokenizers, text_encoders
752
+ ):
753
+ if isinstance(self, TextualInversionLoaderMixin):
754
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
755
+
756
+ text_inputs = tokenizer(
757
+ prompt,
758
+ padding="max_length",
759
+ max_length=tokenizer.model_max_length,
760
+ truncation=True,
761
+ return_tensors="pt",
762
+ )
763
+
764
+ text_input_ids = text_inputs.input_ids
765
+ untruncated_ids = tokenizer(
766
+ prompt, padding="longest", return_tensors="pt"
767
+ ).input_ids
768
+
769
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[
770
+ -1
771
+ ] and not torch.equal(text_input_ids, untruncated_ids):
772
+ removed_text = tokenizer.batch_decode(
773
+ untruncated_ids[:, tokenizer.model_max_length - 1 : -1]
774
+ )
775
+ logger.warning(
776
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
777
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
778
+ )
779
+
780
+ prompt_embeds = text_encoder(
781
+ text_input_ids.to(device),
782
+ output_hidden_states=True,
783
+ )
784
+
785
+ # We are only ALWAYS interested in the pooled output of the final text encoder
786
+ pooled_prompt_embeds = prompt_embeds[0]
787
+ prompt_embeds = prompt_embeds.hidden_states[-2]
788
+
789
+ prompt_embeds_list.append(prompt_embeds)
790
+
791
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
792
+
793
+ # get unconditional embeddings for classifier free guidance
794
+ zero_out_negative_prompt = (
795
+ negative_prompt is None and self.config.force_zeros_for_empty_prompt
796
+ )
797
+ if (
798
+ do_classifier_free_guidance
799
+ and negative_prompt_embeds is None
800
+ and zero_out_negative_prompt
801
+ ):
802
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
803
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
804
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
805
+ negative_prompt = negative_prompt or ""
806
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
807
+
808
+ uncond_tokens: List[str]
809
+ if prompt is not None and type(prompt) is not type(negative_prompt):
810
+ raise TypeError(
811
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
812
+ f" {type(prompt)}."
813
+ )
814
+ elif isinstance(negative_prompt, str):
815
+ uncond_tokens = [negative_prompt, negative_prompt_2]
816
+ elif batch_size != len(negative_prompt):
817
+ raise ValueError(
818
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
819
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
820
+ " the batch size of `prompt`."
821
+ )
822
+ else:
823
+ uncond_tokens = [negative_prompt, negative_prompt_2]
824
+
825
+ negative_prompt_embeds_list = []
826
+ for negative_prompt, tokenizer, text_encoder in zip(
827
+ uncond_tokens, tokenizers, text_encoders
828
+ ):
829
+ if isinstance(self, TextualInversionLoaderMixin):
830
+ negative_prompt = self.maybe_convert_prompt(
831
+ negative_prompt, tokenizer
832
+ )
833
+
834
+ max_length = prompt_embeds.shape[1]
835
+ uncond_input = tokenizer(
836
+ negative_prompt,
837
+ padding="max_length",
838
+ max_length=max_length,
839
+ truncation=True,
840
+ return_tensors="pt",
841
+ )
842
+
843
+ negative_prompt_embeds = text_encoder(
844
+ uncond_input.input_ids.to(device),
845
+ output_hidden_states=True,
846
+ )
847
+ # We are only ALWAYS interested in the pooled output of the final text encoder
848
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
849
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
850
+
851
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
852
+
853
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
854
+
855
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
856
+ bs_embed, seq_len, _ = prompt_embeds.shape
857
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
858
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
859
+ prompt_embeds = prompt_embeds.view(
860
+ bs_embed * num_images_per_prompt, seq_len, -1
861
+ )
862
+
863
+ if do_classifier_free_guidance:
864
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
865
+ seq_len = negative_prompt_embeds.shape[1]
866
+ negative_prompt_embeds = negative_prompt_embeds.to(
867
+ dtype=self.text_encoder_2.dtype, device=device
868
+ )
869
+ negative_prompt_embeds = negative_prompt_embeds.repeat(
870
+ 1, num_images_per_prompt, 1
871
+ )
872
+ negative_prompt_embeds = negative_prompt_embeds.view(
873
+ batch_size * num_images_per_prompt, seq_len, -1
874
+ )
875
+
876
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(
877
+ 1, num_images_per_prompt
878
+ ).view(bs_embed * num_images_per_prompt, -1)
879
+ if do_classifier_free_guidance:
880
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(
881
+ 1, num_images_per_prompt
882
+ ).view(bs_embed * num_images_per_prompt, -1)
883
+
884
+ return (
885
+ prompt_embeds,
886
+ negative_prompt_embeds,
887
+ pooled_prompt_embeds,
888
+ negative_pooled_prompt_embeds,
889
+ )
890
+
891
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
892
+ def prepare_extra_step_kwargs(self, generator, eta):
893
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
894
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
895
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
896
+ # and should be between [0, 1]
897
+
898
+ accepts_eta = "eta" in set(
899
+ inspect.signature(self.scheduler.step).parameters.keys()
900
+ )
901
+ extra_step_kwargs = {}
902
+ if accepts_eta:
903
+ extra_step_kwargs["eta"] = eta
904
+
905
+ # check if the scheduler accepts generator
906
+ accepts_generator = "generator" in set(
907
+ inspect.signature(self.scheduler.step).parameters.keys()
908
+ )
909
+ if accepts_generator:
910
+ extra_step_kwargs["generator"] = generator
911
+ return extra_step_kwargs
912
+
913
+ def check_inputs(
914
+ self,
915
+ prompt,
916
+ prompt_2,
917
+ height,
918
+ width,
919
+ callback_steps,
920
+ negative_prompt=None,
921
+ negative_prompt_2=None,
922
+ prompt_embeds=None,
923
+ negative_prompt_embeds=None,
924
+ pooled_prompt_embeds=None,
925
+ negative_pooled_prompt_embeds=None,
926
+ ):
927
+ if height % 8 != 0 or width % 8 != 0:
928
+ raise ValueError(
929
+ f"`height` and `width` have to be divisible by 8 but are {height} and {width}."
930
+ )
931
+
932
+ if (callback_steps is None) or (
933
+ callback_steps is not None
934
+ and (not isinstance(callback_steps, int) or callback_steps <= 0)
935
+ ):
936
+ raise ValueError(
937
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
938
+ f" {type(callback_steps)}."
939
+ )
940
+
941
+ if prompt is not None and prompt_embeds is not None:
942
+ raise ValueError(
943
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
944
+ " only forward one of the two."
945
+ )
946
+ elif prompt_2 is not None and prompt_embeds is not None:
947
+ raise ValueError(
948
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
949
+ " only forward one of the two."
950
+ )
951
+ elif prompt is None and prompt_embeds is None:
952
+ raise ValueError(
953
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
954
+ )
955
+ elif prompt is not None and (
956
+ not isinstance(prompt, str) and not isinstance(prompt, list)
957
+ ):
958
+ raise ValueError(
959
+ f"`prompt` has to be of type `str` or `list` but is {type(prompt)}"
960
+ )
961
+ elif prompt_2 is not None and (
962
+ not isinstance(prompt_2, str) and not isinstance(prompt_2, list)
963
+ ):
964
+ raise ValueError(
965
+ f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}"
966
+ )
967
+
968
+ if negative_prompt is not None and negative_prompt_embeds is not None:
969
+ raise ValueError(
970
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
971
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
972
+ )
973
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
974
+ raise ValueError(
975
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
976
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
977
+ )
978
+
979
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
980
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
981
+ raise ValueError(
982
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
983
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
984
+ f" {negative_prompt_embeds.shape}."
985
+ )
986
+
987
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
988
+ raise ValueError(
989
+ "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
990
+ )
991
+
992
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
993
+ raise ValueError(
994
+ "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
995
+ )
996
+
997
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
998
+ def prepare_latents(
999
+ self,
1000
+ batch_size,
1001
+ num_channels_latents,
1002
+ height,
1003
+ width,
1004
+ dtype,
1005
+ device,
1006
+ generator,
1007
+ latents=None,
1008
+ ):
1009
+ shape = (
1010
+ batch_size,
1011
+ num_channels_latents,
1012
+ height // self.vae_scale_factor,
1013
+ width // self.vae_scale_factor,
1014
+ )
1015
+ if isinstance(generator, list) and len(generator) != batch_size:
1016
+ raise ValueError(
1017
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
1018
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
1019
+ )
1020
+
1021
+ if latents is None:
1022
+ latents = randn_tensor(
1023
+ shape, generator=generator, device=device, dtype=dtype
1024
+ )
1025
+ else:
1026
+ latents = latents.to(device)
1027
+
1028
+ # scale the initial noise by the standard deviation required by the scheduler
1029
+ latents = latents * self.scheduler.init_noise_sigma
1030
+ return latents
1031
+
1032
+ def _get_add_time_ids(
1033
+ self, original_size, crops_coords_top_left, target_size, dtype
1034
+ ):
1035
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
1036
+
1037
+ passed_add_embed_dim = (
1038
+ self.unet.config.addition_time_embed_dim * len(add_time_ids)
1039
+ + self.text_encoder_2.config.projection_dim
1040
+ )
1041
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
1042
+
1043
+ if expected_add_embed_dim != passed_add_embed_dim:
1044
+ raise ValueError(
1045
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
1046
+ )
1047
+
1048
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
1049
+ return add_time_ids
1050
+
1051
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
1052
+ def upcast_vae(self):
1053
+ dtype = self.vae.dtype
1054
+ self.vae.to(dtype=torch.float32)
1055
+ use_torch_2_0_or_xformers = isinstance(
1056
+ self.vae.decoder.mid_block.attentions[0].processor,
1057
+ (
1058
+ AttnProcessor2_0,
1059
+ XFormersAttnProcessor,
1060
+ LoRAXFormersAttnProcessor,
1061
+ LoRAAttnProcessor2_0,
1062
+ ),
1063
+ )
1064
+ # if xformers or torch_2_0 is used attention block does not need
1065
+ # to be in float32 which can save lots of memory
1066
+ if use_torch_2_0_or_xformers:
1067
+ self.vae.post_quant_conv.to(dtype)
1068
+ self.vae.decoder.conv_in.to(dtype)
1069
+ self.vae.decoder.mid_block.to(dtype)
1070
+
1071
+ @torch.no_grad()
1072
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
1073
+ def __call__(
1074
+ self,
1075
+ prompt: str = None,
1076
+ prompt_2: Optional[str] = None,
1077
+ height: Optional[int] = None,
1078
+ width: Optional[int] = None,
1079
+ num_inference_steps: int = 50,
1080
+ denoising_end: Optional[float] = None,
1081
+ guidance_scale: float = 5.0,
1082
+ negative_prompt: Optional[str] = None,
1083
+ negative_prompt_2: Optional[str] = None,
1084
+ num_images_per_prompt: Optional[int] = 1,
1085
+ eta: float = 0.0,
1086
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
1087
+ latents: Optional[torch.FloatTensor] = None,
1088
+ prompt_embeds: Optional[torch.FloatTensor] = None,
1089
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
1090
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
1091
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
1092
+ output_type: Optional[str] = "pil",
1093
+ return_dict: bool = True,
1094
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
1095
+ callback_steps: int = 1,
1096
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
1097
+ guidance_rescale: float = 0.0,
1098
+ original_size: Optional[Tuple[int, int]] = None,
1099
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
1100
+ target_size: Optional[Tuple[int, int]] = None,
1101
+ ):
1102
+ r"""
1103
+ Function invoked when calling the pipeline for generation.
1104
+
1105
+ Args:
1106
+ prompt (`str`):
1107
+ The prompt to guide the image generation. If not defined, one has to pass `prompt_embeds`.
1108
+ instead.
1109
+ prompt_2 (`str`):
1110
+ The prompt to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
1111
+ used in both text-encoders
1112
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
1113
+ The height in pixels of the generated image.
1114
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
1115
+ The width in pixels of the generated image.
1116
+ num_inference_steps (`int`, *optional*, defaults to 50):
1117
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
1118
+ expense of slower inference.
1119
+ denoising_end (`float`, *optional*):
1120
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
1121
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
1122
+ still retain a substantial amount of noise as determined by the discrete timesteps selected by the
1123
+ scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
1124
+ "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
1125
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
1126
+ guidance_scale (`float`, *optional*, defaults to 5.0):
1127
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
1128
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
1129
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
1130
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
1131
+ usually at the expense of lower image quality.
1132
+ negative_prompt (`str`):
1133
+ The prompt not to guide the image generation. If not defined, one has to pass
1134
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
1135
+ less than `1`).
1136
+ negative_prompt_2 (`str`):
1137
+ The prompt not to guide the image generation to be sent to `tokenizer_2` and
1138
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
1139
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
1140
+ The number of images to generate per prompt.
1141
+ eta (`float`, *optional*, defaults to 0.0):
1142
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
1143
+ [`schedulers.DDIMScheduler`], will be ignored for others.
1144
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
1145
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
1146
+ to make generation deterministic.
1147
+ latents (`torch.FloatTensor`, *optional*):
1148
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
1149
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
1150
+ tensor will ge generated by sampling using the supplied random `generator`.
1151
+ prompt_embeds (`torch.FloatTensor`, *optional*):
1152
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
1153
+ provided, text embeddings will be generated from `prompt` input argument.
1154
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
1155
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
1156
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
1157
+ argument.
1158
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
1159
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
1160
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
1161
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
1162
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
1163
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
1164
+ input argument.
1165
+ output_type (`str`, *optional*, defaults to `"pil"`):
1166
+ The output format of the generate image. Choose between
1167
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
1168
+ return_dict (`bool`, *optional*, defaults to `True`):
1169
+ Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
1170
+ of a plain tuple.
1171
+ callback (`Callable`, *optional*):
1172
+ A function that will be called every `callback_steps` steps during inference. The function will be
1173
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
1174
+ callback_steps (`int`, *optional*, defaults to 1):
1175
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
1176
+ called at every step.
1177
+ cross_attention_kwargs (`dict`, *optional*):
1178
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
1179
+ `self.processor` in
1180
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
1181
+ guidance_rescale (`float`, *optional*, defaults to 0.7):
1182
+ Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
1183
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
1184
+ [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
1185
+ Guidance rescale factor should fix overexposure when using zero terminal SNR.
1186
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
1187
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
1188
+ `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
1189
+ explained in section 2.2 of
1190
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
1191
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
1192
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
1193
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
1194
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
1195
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
1196
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
1197
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
1198
+ not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
1199
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
1200
+
1201
+ Examples:
1202
+
1203
+ Returns:
1204
+ [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
1205
+ [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
1206
+ `tuple`. When returning a tuple, the first element is a list with the generated images.
1207
+ """
1208
+ # 0. Default height and width to unet
1209
+ height = height or self.default_sample_size * self.vae_scale_factor
1210
+ width = width or self.default_sample_size * self.vae_scale_factor
1211
+
1212
+ original_size = original_size or (height, width)
1213
+ target_size = target_size or (height, width)
1214
+
1215
+ # 1. Check inputs. Raise error if not correct
1216
+ self.check_inputs(
1217
+ prompt,
1218
+ prompt_2,
1219
+ height,
1220
+ width,
1221
+ callback_steps,
1222
+ negative_prompt,
1223
+ negative_prompt_2,
1224
+ prompt_embeds,
1225
+ negative_prompt_embeds,
1226
+ pooled_prompt_embeds,
1227
+ negative_pooled_prompt_embeds,
1228
+ )
1229
+
1230
+ # 2. Define call parameters
1231
+ if prompt is not None and isinstance(prompt, str):
1232
+ batch_size = 1
1233
+ elif prompt is not None and isinstance(prompt, list):
1234
+ batch_size = len(prompt)
1235
+ else:
1236
+ batch_size = prompt_embeds.shape[0]
1237
+
1238
+ device = self._execution_device
1239
+
1240
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
1241
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
1242
+ # corresponds to doing no classifier free guidance.
1243
+ do_classifier_free_guidance = guidance_scale > 1.0
1244
+
1245
+ # 3. Encode input prompt
1246
+ (
1247
+ cross_attention_kwargs.get("scale", None)
1248
+ if cross_attention_kwargs is not None
1249
+ else None
1250
+ )
1251
+
1252
+ negative_prompt = negative_prompt if negative_prompt is not None else ""
1253
+
1254
+ (
1255
+ prompt_embeds,
1256
+ negative_prompt_embeds,
1257
+ pooled_prompt_embeds,
1258
+ negative_pooled_prompt_embeds,
1259
+ ) = get_weighted_text_embeddings_sdxl(
1260
+ pipe=self, prompt=prompt, neg_prompt=negative_prompt
1261
+ )
1262
+
1263
+ # 4. Prepare timesteps
1264
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
1265
+
1266
+ timesteps = self.scheduler.timesteps
1267
+
1268
+ # 5. Prepare latent variables
1269
+ num_channels_latents = self.unet.config.in_channels
1270
+ latents = self.prepare_latents(
1271
+ batch_size * num_images_per_prompt,
1272
+ num_channels_latents,
1273
+ height,
1274
+ width,
1275
+ prompt_embeds.dtype,
1276
+ device,
1277
+ generator,
1278
+ latents,
1279
+ )
1280
+
1281
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
1282
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
1283
+
1284
+ # 7. Prepare added time ids & embeddings
1285
+ add_text_embeds = pooled_prompt_embeds
1286
+ add_time_ids = self._get_add_time_ids(
1287
+ original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
1288
+ )
1289
+
1290
+ if do_classifier_free_guidance:
1291
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
1292
+ add_text_embeds = torch.cat(
1293
+ [negative_pooled_prompt_embeds, add_text_embeds], dim=0
1294
+ )
1295
+ add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
1296
+
1297
+ prompt_embeds = prompt_embeds.to(device)
1298
+ add_text_embeds = add_text_embeds.to(device)
1299
+ add_time_ids = add_time_ids.to(device).repeat(
1300
+ batch_size * num_images_per_prompt, 1
1301
+ )
1302
+
1303
+ # 8. Denoising loop
1304
+ num_warmup_steps = max(
1305
+ len(timesteps) - num_inference_steps * self.scheduler.order, 0
1306
+ )
1307
+
1308
+ # 7.1 Apply denoising_end
1309
+ if (
1310
+ denoising_end is not None
1311
+ and type(denoising_end) == float
1312
+ and denoising_end > 0
1313
+ and denoising_end < 1
1314
+ ):
1315
+ discrete_timestep_cutoff = int(
1316
+ round(
1317
+ self.scheduler.config.num_train_timesteps
1318
+ - (denoising_end * self.scheduler.config.num_train_timesteps)
1319
+ )
1320
+ )
1321
+ num_inference_steps = len(
1322
+ list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))
1323
+ )
1324
+ timesteps = timesteps[:num_inference_steps]
1325
+
1326
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
1327
+ for i, t in enumerate(timesteps):
1328
+ # expand the latents if we are doing classifier free guidance
1329
+ latent_model_input = (
1330
+ torch.cat([latents] * 2) if do_classifier_free_guidance else latents
1331
+ )
1332
+
1333
+ latent_model_input = self.scheduler.scale_model_input(
1334
+ latent_model_input, t
1335
+ )
1336
+
1337
+ # predict the noise residual
1338
+ added_cond_kwargs = {
1339
+ "text_embeds": add_text_embeds,
1340
+ "time_ids": add_time_ids,
1341
+ }
1342
+ noise_pred = self.unet(
1343
+ latent_model_input,
1344
+ t,
1345
+ encoder_hidden_states=prompt_embeds,
1346
+ cross_attention_kwargs=cross_attention_kwargs,
1347
+ added_cond_kwargs=added_cond_kwargs,
1348
+ return_dict=False,
1349
+ )[0]
1350
+
1351
+ # perform guidance
1352
+ if do_classifier_free_guidance:
1353
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
1354
+ noise_pred = noise_pred_uncond + guidance_scale * (
1355
+ noise_pred_text - noise_pred_uncond
1356
+ )
1357
+
1358
+ if do_classifier_free_guidance and guidance_rescale > 0.0:
1359
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
1360
+ noise_pred = rescale_noise_cfg(
1361
+ noise_pred, noise_pred_text, guidance_rescale=guidance_rescale
1362
+ )
1363
+
1364
+ # compute the previous noisy sample x_t -> x_t-1
1365
+ latents = self.scheduler.step(
1366
+ noise_pred, t, latents, **extra_step_kwargs, return_dict=False
1367
+ )[0]
1368
+
1369
+ # call the callback, if provided
1370
+ if i == len(timesteps) - 1 or (
1371
+ (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
1372
+ ):
1373
+ progress_bar.update()
1374
+ if callback is not None and i % callback_steps == 0:
1375
+ callback(i, t, latents)
1376
+
1377
+ # make sure the VAE is in float32 mode, as it overflows in float16
1378
+ if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
1379
+ self.upcast_vae()
1380
+ latents = latents.to(
1381
+ next(iter(self.vae.post_quant_conv.parameters())).dtype
1382
+ )
1383
+
1384
+ if not output_type == "latent":
1385
+ image = self.vae.decode(
1386
+ latents / self.vae.config.scaling_factor, return_dict=False
1387
+ )[0]
1388
+ else:
1389
+ image = latents
1390
+ return StableDiffusionXLPipelineOutput(images=image)
1391
+
1392
+ # apply watermark if available
1393
+ if self.watermark is not None:
1394
+ image = self.watermark.apply_watermark(image)
1395
+
1396
+ image = self.image_processor.postprocess(image, output_type=output_type)
1397
+
1398
+ # Offload last model to CPU
1399
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
1400
+ self.final_offload_hook.offload()
1401
+
1402
+ if not return_dict:
1403
+ return (image,)
1404
+
1405
+ return StableDiffusionXLPipelineOutput(images=image)
1406
+
1407
+ # Overrride to properly handle the loading and unloading of the additional text encoder.
1408
+ def load_lora_weights(
1409
+ self,
1410
+ pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
1411
+ **kwargs,
1412
+ ):
1413
+ # We could have accessed the unet config from `lora_state_dict()` too. We pass
1414
+ # it here explicitly to be able to tell that it's coming from an SDXL
1415
+ # pipeline.
1416
+ state_dict, network_alphas = self.lora_state_dict(
1417
+ pretrained_model_name_or_path_or_dict,
1418
+ unet_config=self.unet.config,
1419
+ **kwargs,
1420
+ )
1421
+ self.load_lora_into_unet(
1422
+ state_dict, network_alphas=network_alphas, unet=self.unet
1423
+ )
1424
+
1425
+ text_encoder_state_dict = {
1426
+ k: v for k, v in state_dict.items() if "text_encoder." in k
1427
+ }
1428
+ if len(text_encoder_state_dict) > 0:
1429
+ self.load_lora_into_text_encoder(
1430
+ text_encoder_state_dict,
1431
+ network_alphas=network_alphas,
1432
+ text_encoder=self.text_encoder,
1433
+ prefix="text_encoder",
1434
+ lora_scale=self.lora_scale,
1435
+ )
1436
+
1437
+ text_encoder_2_state_dict = {
1438
+ k: v for k, v in state_dict.items() if "text_encoder_2." in k
1439
+ }
1440
+ if len(text_encoder_2_state_dict) > 0:
1441
+ self.load_lora_into_text_encoder(
1442
+ text_encoder_2_state_dict,
1443
+ network_alphas=network_alphas,
1444
+ text_encoder=self.text_encoder_2,
1445
+ prefix="text_encoder_2",
1446
+ lora_scale=self.lora_scale,
1447
+ )
1448
+
1449
+ @classmethod
1450
+ def save_lora_weights(
1451
+ self,
1452
+ save_directory: Union[str, os.PathLike],
1453
+ unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
1454
+ text_encoder_lora_layers: Dict[
1455
+ str, Union[torch.nn.Module, torch.Tensor]
1456
+ ] = None,
1457
+ text_encoder_2_lora_layers: Dict[
1458
+ str, Union[torch.nn.Module, torch.Tensor]
1459
+ ] = None,
1460
+ is_main_process: bool = True,
1461
+ weight_name: str = None,
1462
+ save_function: Callable = None,
1463
+ safe_serialization: bool = False,
1464
+ ):
1465
+ state_dict = {}
1466
+
1467
+ def pack_weights(layers, prefix):
1468
+ layers_weights = (
1469
+ layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
1470
+ )
1471
+ layers_state_dict = {
1472
+ f"{prefix}.{module_name}": param
1473
+ for module_name, param in layers_weights.items()
1474
+ }
1475
+ return layers_state_dict
1476
+
1477
+ state_dict.update(pack_weights(unet_lora_layers, "unet"))
1478
+
1479
+ if text_encoder_lora_layers and text_encoder_2_lora_layers:
1480
+ state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
1481
+ state_dict.update(
1482
+ pack_weights(text_encoder_2_lora_layers, "text_encoder_2")
1483
+ )
1484
+
1485
+ self.write_lora_layers(
1486
+ state_dict=state_dict,
1487
+ save_directory=save_directory,
1488
+ is_main_process=is_main_process,
1489
+ weight_name=weight_name,
1490
+ save_function=save_function,
1491
+ safe_serialization=safe_serialization,
1492
+ )
1493
+
1494
+ def _remove_text_encoder_monkey_patch(self):
1495
+ self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder)
1496
+ self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2)
requirements.txt CHANGED
@@ -1,7 +1,8 @@
1
  accelerate==0.21.0
2
- diffusers==0.19.3
3
  gradio==3.40.1
4
  invisible-watermark==0.2.0
5
  Pillow==10.0.0
6
  torch==2.0.1
7
- transformers==4.31.0
 
 
1
  accelerate==0.21.0
2
+ diffusers==0.20.0
3
  gradio==3.40.1
4
  invisible-watermark==0.2.0
5
  Pillow==10.0.0
6
  torch==2.0.1
7
+ transformers==4.31.0
8
+ toml==0.10.2
style.css CHANGED
@@ -3,6 +3,11 @@ h1 {
3
  font-size: 10vw; /* relative to the viewport width */
4
  }
5
 
 
 
 
 
 
6
  #duplicate-button {
7
  margin: auto;
8
  color: #fff;
@@ -23,3 +28,33 @@ h1 {
23
  padding-top: 1rem;
24
  }
25
  }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  font-size: 10vw; /* relative to the viewport width */
4
  }
5
 
6
+ h2 {
7
+ text-align: center;
8
+ font-size: 10vw; /* relative to the viewport width */
9
+ }
10
+
11
  #duplicate-button {
12
  margin: auto;
13
  color: #fff;
 
28
  padding-top: 1rem;
29
  }
30
  }
31
+
32
+ #gallery .grid-wrap{
33
+ min-height: 25%;
34
+ }
35
+
36
+ #title-container {
37
+ display: flex;
38
+ justify-content: center;
39
+ align-items: center;
40
+ height: 100vh; /* Adjust this value to position the title vertically */
41
+ }
42
+
43
+ #title {
44
+ font-size: 3em;
45
+ text-align: center;
46
+ color: #333;
47
+ font-family: 'Helvetica Neue', sans-serif;
48
+ text-transform: uppercase;
49
+ background: transparent;
50
+ }
51
+
52
+ #title span {
53
+ background: -webkit-linear-gradient(45deg, #4EACEF, #28b485);
54
+ -webkit-background-clip: text;
55
+ -webkit-text-fill-color: transparent;
56
+ }
57
+
58
+ #subtitle {
59
+ text-align: center;
60
+ }
utils.py ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ def is_google_colab():
2
+ try:
3
+ import google.colab
4
+
5
+ return True
6
+ except:
7
+ return False