Create new file
Browse files- pipeline.py +1137 -0
pipeline.py
ADDED
@@ -0,0 +1,1137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import inspect
|
2 |
+
import re
|
3 |
+
from typing import Callable, List, Optional, Union
|
4 |
+
import PIL
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
|
8 |
+
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
|
9 |
+
|
10 |
+
from diffusers.configuration_utils import FrozenDict
|
11 |
+
from diffusers.models import AutoencoderKL, UNet2DConditionModel
|
12 |
+
from diffusers.pipeline_utils import DiffusionPipeline
|
13 |
+
from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
|
14 |
+
from diffusers.utils import deprecate, logging
|
15 |
+
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
|
16 |
+
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
|
17 |
+
|
18 |
+
|
19 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
20 |
+
|
21 |
+
re_attention = re.compile(r"""
|
22 |
+
\\\(|
|
23 |
+
\\\)|
|
24 |
+
\\\[|
|
25 |
+
\\]|
|
26 |
+
\\\\|
|
27 |
+
\\|
|
28 |
+
\(|
|
29 |
+
\[|
|
30 |
+
:([+-]?[.\d]+)\)|
|
31 |
+
\)|
|
32 |
+
]|
|
33 |
+
[^\\()\[\]:]+|
|
34 |
+
:
|
35 |
+
""", re.X)
|
36 |
+
|
37 |
+
|
38 |
+
def parse_prompt_attention(text):
|
39 |
+
"""
|
40 |
+
Parses a string with attention tokens and returns a list of pairs: text and its assoicated weight.
|
41 |
+
Accepted tokens are:
|
42 |
+
(abc) - increases attention to abc by a multiplier of 1.1
|
43 |
+
(abc:3.12) - increases attention to abc by a multiplier of 3.12
|
44 |
+
[abc] - decreases attention to abc by a multiplier of 1.1
|
45 |
+
\( - literal character '('
|
46 |
+
\[ - literal character '['
|
47 |
+
\) - literal character ')'
|
48 |
+
\] - literal character ']'
|
49 |
+
\\ - literal character '\'
|
50 |
+
anything else - just text
|
51 |
+
>>> parse_prompt_attention('normal text')
|
52 |
+
[['normal text', 1.0]]
|
53 |
+
>>> parse_prompt_attention('an (important) word')
|
54 |
+
[['an ', 1.0], ['important', 1.1], [' word', 1.0]]
|
55 |
+
>>> parse_prompt_attention('(unbalanced')
|
56 |
+
[['unbalanced', 1.1]]
|
57 |
+
>>> parse_prompt_attention('\(literal\]')
|
58 |
+
[['(literal]', 1.0]]
|
59 |
+
>>> parse_prompt_attention('(unnecessary)(parens)')
|
60 |
+
[['unnecessaryparens', 1.1]]
|
61 |
+
>>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).')
|
62 |
+
[['a ', 1.0],
|
63 |
+
['house', 1.5730000000000004],
|
64 |
+
[' ', 1.1],
|
65 |
+
['on', 1.0],
|
66 |
+
[' a ', 1.1],
|
67 |
+
['hill', 0.55],
|
68 |
+
[', sun, ', 1.1],
|
69 |
+
['sky', 1.4641000000000006],
|
70 |
+
['.', 1.1]]
|
71 |
+
"""
|
72 |
+
|
73 |
+
res = []
|
74 |
+
round_brackets = []
|
75 |
+
square_brackets = []
|
76 |
+
|
77 |
+
round_bracket_multiplier = 1.1
|
78 |
+
square_bracket_multiplier = 1 / 1.1
|
79 |
+
|
80 |
+
def multiply_range(start_position, multiplier):
|
81 |
+
for p in range(start_position, len(res)):
|
82 |
+
res[p][1] *= multiplier
|
83 |
+
|
84 |
+
for m in re_attention.finditer(text):
|
85 |
+
text = m.group(0)
|
86 |
+
weight = m.group(1)
|
87 |
+
|
88 |
+
if text.startswith('\\'):
|
89 |
+
res.append([text[1:], 1.0])
|
90 |
+
elif text == '(':
|
91 |
+
round_brackets.append(len(res))
|
92 |
+
elif text == '[':
|
93 |
+
square_brackets.append(len(res))
|
94 |
+
elif weight is not None and len(round_brackets) > 0:
|
95 |
+
multiply_range(round_brackets.pop(), float(weight))
|
96 |
+
elif text == ')' and len(round_brackets) > 0:
|
97 |
+
multiply_range(round_brackets.pop(), round_bracket_multiplier)
|
98 |
+
elif text == ']' and len(square_brackets) > 0:
|
99 |
+
multiply_range(square_brackets.pop(), square_bracket_multiplier)
|
100 |
+
else:
|
101 |
+
res.append([text, 1.0])
|
102 |
+
|
103 |
+
for pos in round_brackets:
|
104 |
+
multiply_range(pos, round_bracket_multiplier)
|
105 |
+
|
106 |
+
for pos in square_brackets:
|
107 |
+
multiply_range(pos, square_bracket_multiplier)
|
108 |
+
|
109 |
+
if len(res) == 0:
|
110 |
+
res = [["", 1.0]]
|
111 |
+
|
112 |
+
# merge runs of identical weights
|
113 |
+
i = 0
|
114 |
+
while i + 1 < len(res):
|
115 |
+
if res[i][1] == res[i + 1][1]:
|
116 |
+
res[i][0] += res[i + 1][0]
|
117 |
+
res.pop(i + 1)
|
118 |
+
else:
|
119 |
+
i += 1
|
120 |
+
|
121 |
+
return res
|
122 |
+
|
123 |
+
|
124 |
+
def get_prompts_with_weights(
|
125 |
+
pipe: DiffusionPipeline,
|
126 |
+
prompt: List[str],
|
127 |
+
max_length: int
|
128 |
+
):
|
129 |
+
r"""
|
130 |
+
Tokenize a list of prompts and return its tokens with weights of each token.
|
131 |
+
|
132 |
+
No padding, starting or ending token is included.
|
133 |
+
"""
|
134 |
+
tokens = []
|
135 |
+
weights = []
|
136 |
+
for text in prompt:
|
137 |
+
texts_and_weights = parse_prompt_attention(text)
|
138 |
+
text_token = []
|
139 |
+
text_weight = []
|
140 |
+
for word, weight in texts_and_weights:
|
141 |
+
# tokenize and discard the starting and the ending token
|
142 |
+
token = pipe.tokenizer(word).input_ids[1:-1]
|
143 |
+
text_token += token
|
144 |
+
|
145 |
+
# copy the weight by length of token
|
146 |
+
text_weight += [weight] * len(token)
|
147 |
+
|
148 |
+
# stop if the text is too long (longer than truncation limit)
|
149 |
+
if len(text_token) > max_length:
|
150 |
+
break
|
151 |
+
|
152 |
+
# truncate
|
153 |
+
if len(text_token) > max_length:
|
154 |
+
text_token = text_token[:max_length]
|
155 |
+
text_weight = text_weight[:max_length]
|
156 |
+
|
157 |
+
tokens.append(text_token)
|
158 |
+
weights.append(text_weight)
|
159 |
+
return tokens, weights
|
160 |
+
|
161 |
+
|
162 |
+
def pad_tokens_and_weights(tokens, weights, max_length, bos, eos,
|
163 |
+
no_boseos_middle=True,
|
164 |
+
chunk_length=77):
|
165 |
+
r"""
|
166 |
+
Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length.
|
167 |
+
"""
|
168 |
+
max_embeddings_multiples = (max_length - 2) // (chunk_length - 2)
|
169 |
+
weights_length = max_length if no_boseos_middle else max_embeddings_multiples * chunk_length
|
170 |
+
for i in range(len(tokens)):
|
171 |
+
tokens[i] = [bos] + tokens[i] + [eos] * (max_length - 1 - len(tokens[i]))
|
172 |
+
if no_boseos_middle:
|
173 |
+
weights[i] = [1.] + weights[i] + [1.] * (max_length - 1 - len(weights[i]))
|
174 |
+
else:
|
175 |
+
w = []
|
176 |
+
if len(weights[i]) == 0:
|
177 |
+
w = [1.] * weights_length
|
178 |
+
else:
|
179 |
+
for j in range((len(weights[i]) - 1) // chunk_length + 1):
|
180 |
+
w.append(1.) # weight for starting token in this chunk
|
181 |
+
w += weights[i][j * chunk_length: min(len(weights[i]), (j + 1) * chunk_length)]
|
182 |
+
w.append(1.) # weight for ending token in this chunk
|
183 |
+
w += [1.] * (weights_length - len(w))
|
184 |
+
weights[i] = w[:]
|
185 |
+
|
186 |
+
return tokens, weights
|
187 |
+
|
188 |
+
|
189 |
+
def get_unweighted_text_embeddings(
|
190 |
+
pipe: DiffusionPipeline,
|
191 |
+
text_input: torch.Tensor,
|
192 |
+
chunk_length: int,
|
193 |
+
no_boseos_middle: Optional[bool] = True
|
194 |
+
):
|
195 |
+
"""
|
196 |
+
When the length of tokens is a multiple of the capacity of the text encoder,
|
197 |
+
it should be split into chunks and sent to the text encoder individually.
|
198 |
+
"""
|
199 |
+
max_embeddings_multiples = (text_input.shape[1] - 2) // (chunk_length - 2)
|
200 |
+
if max_embeddings_multiples > 1:
|
201 |
+
text_embeddings = []
|
202 |
+
for i in range(max_embeddings_multiples):
|
203 |
+
# extract the i-th chunk
|
204 |
+
text_input_chunk = text_input[:, i * (chunk_length - 2):(i + 1) * (chunk_length - 2) + 2].clone()
|
205 |
+
|
206 |
+
# cover the head and the tail by the starting and the ending tokens
|
207 |
+
text_input_chunk[:, 0] = text_input[0, 0]
|
208 |
+
text_input_chunk[:, -1] = text_input[0, -1]
|
209 |
+
text_embedding = pipe.text_encoder(text_input_chunk)[0]
|
210 |
+
|
211 |
+
if no_boseos_middle:
|
212 |
+
if i == 0:
|
213 |
+
# discard the ending token
|
214 |
+
text_embedding = text_embedding[:, :-1]
|
215 |
+
elif i == max_embeddings_multiples - 1:
|
216 |
+
# discard the starting token
|
217 |
+
text_embedding = text_embedding[:, 1:]
|
218 |
+
else:
|
219 |
+
# discard both starting and ending tokens
|
220 |
+
text_embedding = text_embedding[:, 1:-1]
|
221 |
+
|
222 |
+
text_embeddings.append(text_embedding)
|
223 |
+
text_embeddings = torch.concat(text_embeddings, axis=1)
|
224 |
+
else:
|
225 |
+
text_embeddings = pipe.text_encoder(text_input)[0]
|
226 |
+
return text_embeddings
|
227 |
+
|
228 |
+
|
229 |
+
def get_weighted_text_embeddings(
|
230 |
+
pipe: DiffusionPipeline,
|
231 |
+
prompt: Union[str, List[str]],
|
232 |
+
uncond_prompt: Optional[Union[str, List[str]]] = None,
|
233 |
+
max_embeddings_multiples: Optional[int] = 1,
|
234 |
+
no_boseos_middle: Optional[bool] = False,
|
235 |
+
skip_parsing: Optional[bool] = False,
|
236 |
+
skip_weighting: Optional[bool] = False,
|
237 |
+
**kwargs
|
238 |
+
):
|
239 |
+
r"""
|
240 |
+
Prompts can be assigned with local weights using brackets. For example,
|
241 |
+
prompt 'A (very beautiful) masterpiece' highlights the words 'very beautiful',
|
242 |
+
and the embedding tokens corresponding to the words get multipled by a constant, 1.1.
|
243 |
+
|
244 |
+
Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the origional mean.
|
245 |
+
|
246 |
+
Args:
|
247 |
+
pipe (`DiffusionPipeline`):
|
248 |
+
Pipe to provide access to the tokenizer and the text encoder.
|
249 |
+
prompt (`str` or `List[str]`):
|
250 |
+
The prompt or prompts to guide the image generation.
|
251 |
+
uncond_prompt (`str` or `List[str]`):
|
252 |
+
The unconditional prompt or prompts for guide the image generation. If unconditional prompt
|
253 |
+
is provided, the embeddings of prompt and uncond_prompt are concatenated.
|
254 |
+
max_embeddings_multiples (`int`, *optional*, defaults to `1`):
|
255 |
+
The max multiple length of prompt embeddings compared to the max output length of text encoder.
|
256 |
+
no_boseos_middle (`bool`, *optional*, defaults to `False`):
|
257 |
+
If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and
|
258 |
+
ending token in each of the chunk in the middle.
|
259 |
+
skip_parsing (`bool`, *optional*, defaults to `False`):
|
260 |
+
Skip the parsing of brackets.
|
261 |
+
skip_weighting (`bool`, *optional*, defaults to `False`):
|
262 |
+
Skip the weighting. When the parsing is skipped, it is forced True.
|
263 |
+
"""
|
264 |
+
max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2
|
265 |
+
if isinstance(prompt, str):
|
266 |
+
prompt = [prompt]
|
267 |
+
|
268 |
+
if not skip_parsing:
|
269 |
+
prompt_tokens, prompt_weights = get_prompts_with_weights(pipe, prompt, max_length - 2)
|
270 |
+
if uncond_prompt is not None:
|
271 |
+
if isinstance(uncond_prompt, str):
|
272 |
+
uncond_prompt = [uncond_prompt]
|
273 |
+
uncond_tokens, uncond_weights = get_prompts_with_weights(pipe, uncond_prompt, max_length - 2)
|
274 |
+
else:
|
275 |
+
prompt_tokens = [token[1:-1] for token in
|
276 |
+
pipe.tokenizer(prompt, max_length=max_length, truncation=True).input_ids]
|
277 |
+
prompt_weights = [[1.] * len(token) for token in prompt_tokens]
|
278 |
+
if uncond_prompt is not None:
|
279 |
+
if isinstance(uncond_prompt, str):
|
280 |
+
uncond_prompt = [uncond_prompt]
|
281 |
+
uncond_tokens = [token[1:-1] for token in
|
282 |
+
pipe.tokenizer(uncond_prompt, max_length=max_length, truncation=True).input_ids]
|
283 |
+
uncond_weights = [[1.] * len(token) for token in uncond_tokens]
|
284 |
+
|
285 |
+
# round up the longest length of tokens to a multiple of (model_max_length - 2)
|
286 |
+
max_length = max([len(token) for token in prompt_tokens])
|
287 |
+
if uncond_prompt is not None:
|
288 |
+
max_length = max(max_length, max([len(token) for token in uncond_tokens]))
|
289 |
+
|
290 |
+
max_embeddings_multiples = min(max_embeddings_multiples,
|
291 |
+
(max_length - 1) // (pipe.tokenizer.model_max_length - 2) + 1)
|
292 |
+
max_embeddings_multiples = max(1, max_embeddings_multiples)
|
293 |
+
max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2
|
294 |
+
|
295 |
+
# pad the length of tokens and weights
|
296 |
+
bos = pipe.tokenizer.bos_token_id
|
297 |
+
eos = pipe.tokenizer.eos_token_id
|
298 |
+
prompt_tokens, prompt_weights = pad_tokens_and_weights(prompt_tokens, prompt_weights, max_length, bos, eos,
|
299 |
+
no_boseos_middle=no_boseos_middle,
|
300 |
+
chunk_length=pipe.tokenizer.model_max_length)
|
301 |
+
prompt_tokens = torch.tensor(prompt_tokens, dtype=torch.long, device=pipe.device)
|
302 |
+
if uncond_prompt is not None:
|
303 |
+
uncond_tokens, uncond_weights = pad_tokens_and_weights(uncond_tokens, uncond_weights, max_length, bos, eos,
|
304 |
+
no_boseos_middle=no_boseos_middle,
|
305 |
+
chunk_length=pipe.tokenizer.model_max_length)
|
306 |
+
uncond_tokens = torch.tensor(uncond_tokens, dtype=torch.long, device=pipe.device)
|
307 |
+
|
308 |
+
# get the embeddings
|
309 |
+
text_embeddings = get_unweighted_text_embeddings(pipe, prompt_tokens, pipe.tokenizer.model_max_length,
|
310 |
+
no_boseos_middle=no_boseos_middle)
|
311 |
+
prompt_weights = torch.tensor(prompt_weights, dtype=text_embeddings.dtype, device=pipe.device)
|
312 |
+
if uncond_prompt is not None:
|
313 |
+
uncond_embeddings = get_unweighted_text_embeddings(pipe, uncond_tokens, pipe.tokenizer.model_max_length,
|
314 |
+
no_boseos_middle=no_boseos_middle)
|
315 |
+
uncond_weights = torch.tensor(uncond_weights, dtype=uncond_embeddings.dtype, device=pipe.device)
|
316 |
+
|
317 |
+
# assign weights to the prompts and normalize in the sense of mean
|
318 |
+
# TODO: should we normalize by chunk or in a whole (current implementation)?
|
319 |
+
if (not skip_parsing) and (not skip_weighting):
|
320 |
+
previous_mean = text_embeddings.mean(axis=[-2, -1])
|
321 |
+
text_embeddings *= prompt_weights.unsqueeze(-1)
|
322 |
+
text_embeddings *= previous_mean / text_embeddings.mean(axis=[-2, -1])
|
323 |
+
if uncond_prompt is not None:
|
324 |
+
previous_mean = uncond_embeddings.mean(axis=[-2, -1])
|
325 |
+
uncond_embeddings *= uncond_weights.unsqueeze(-1)
|
326 |
+
uncond_embeddings *= previous_mean / uncond_embeddings.mean(axis=[-2, -1])
|
327 |
+
|
328 |
+
# For classifier free guidance, we need to do two forward passes.
|
329 |
+
# Here we concatenate the unconditional and text embeddings into a single batch
|
330 |
+
# to avoid doing two forward passes
|
331 |
+
if uncond_prompt is not None:
|
332 |
+
text_embeddings = torch.concat([uncond_embeddings, text_embeddings])
|
333 |
+
|
334 |
+
return text_embeddings
|
335 |
+
|
336 |
+
|
337 |
+
def preprocess_image(image):
|
338 |
+
w, h = image.size
|
339 |
+
w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
|
340 |
+
image = image.resize((w, h), resample=PIL.Image.LANCZOS)
|
341 |
+
image = np.array(image).astype(np.float32) / 255.0
|
342 |
+
image = image[None].transpose(0, 3, 1, 2)
|
343 |
+
image = torch.from_numpy(image)
|
344 |
+
return 2.0 * image - 1.0
|
345 |
+
|
346 |
+
|
347 |
+
def preprocess_mask(mask):
|
348 |
+
mask = mask.convert("L")
|
349 |
+
w, h = mask.size
|
350 |
+
w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
|
351 |
+
mask = mask.resize((w // 8, h // 8), resample=PIL.Image.NEAREST)
|
352 |
+
mask = np.array(mask).astype(np.float32) / 255.0
|
353 |
+
mask = np.tile(mask, (4, 1, 1))
|
354 |
+
mask = mask[None].transpose(0, 1, 2, 3) # what does this step do?
|
355 |
+
mask = 1 - mask # repaint white, keep black
|
356 |
+
mask = torch.from_numpy(mask)
|
357 |
+
return mask
|
358 |
+
|
359 |
+
|
360 |
+
class StableDiffusionLongPromptPipeline(DiffusionPipeline):
|
361 |
+
r"""
|
362 |
+
Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing
|
363 |
+
weighting in prompt.
|
364 |
+
|
365 |
+
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
|
366 |
+
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
|
367 |
+
|
368 |
+
Args:
|
369 |
+
vae ([`AutoencoderKL`]):
|
370 |
+
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
|
371 |
+
text_encoder ([`CLIPTextModel`]):
|
372 |
+
Frozen text-encoder. Stable Diffusion uses the text portion of
|
373 |
+
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
|
374 |
+
the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
|
375 |
+
tokenizer (`CLIPTokenizer`):
|
376 |
+
Tokenizer of class
|
377 |
+
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
|
378 |
+
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
|
379 |
+
scheduler ([`SchedulerMixin`]):
|
380 |
+
A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
|
381 |
+
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
|
382 |
+
safety_checker ([`StableDiffusionSafetyChecker`]):
|
383 |
+
Classification module that estimates whether generated images could be considered offensive or harmful.
|
384 |
+
Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
|
385 |
+
feature_extractor ([`CLIPFeatureExtractor`]):
|
386 |
+
Model that extracts features from generated images to be used as inputs for the `safety_checker`.
|
387 |
+
"""
|
388 |
+
|
389 |
+
def __init__(
|
390 |
+
self,
|
391 |
+
vae: AutoencoderKL,
|
392 |
+
text_encoder: CLIPTextModel,
|
393 |
+
tokenizer: CLIPTokenizer,
|
394 |
+
unet: UNet2DConditionModel,
|
395 |
+
scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
|
396 |
+
safety_checker: StableDiffusionSafetyChecker,
|
397 |
+
feature_extractor: CLIPFeatureExtractor,
|
398 |
+
):
|
399 |
+
super().__init__()
|
400 |
+
|
401 |
+
if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
|
402 |
+
deprecation_message = (
|
403 |
+
f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
|
404 |
+
f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
|
405 |
+
"to update the config accordingly as leaving `steps_offset` might led to incorrect results"
|
406 |
+
" in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
|
407 |
+
" it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
|
408 |
+
" file"
|
409 |
+
)
|
410 |
+
deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
|
411 |
+
new_config = dict(scheduler.config)
|
412 |
+
new_config["steps_offset"] = 1
|
413 |
+
scheduler._internal_dict = FrozenDict(new_config)
|
414 |
+
|
415 |
+
if safety_checker is None:
|
416 |
+
logger.warn(
|
417 |
+
f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
|
418 |
+
" that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
|
419 |
+
" results in services or applications open to the public. Both the diffusers team and Hugging Face"
|
420 |
+
" strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
|
421 |
+
" it only for use-cases that involve analyzing network behavior or auditing its results. For more"
|
422 |
+
" information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
|
423 |
+
)
|
424 |
+
|
425 |
+
self.register_modules(
|
426 |
+
vae=vae,
|
427 |
+
text_encoder=text_encoder,
|
428 |
+
tokenizer=tokenizer,
|
429 |
+
unet=unet,
|
430 |
+
scheduler=scheduler,
|
431 |
+
safety_checker=safety_checker,
|
432 |
+
feature_extractor=feature_extractor,
|
433 |
+
)
|
434 |
+
|
435 |
+
def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
|
436 |
+
r"""
|
437 |
+
Enable sliced attention computation.
|
438 |
+
|
439 |
+
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
|
440 |
+
in several steps. This is useful to save some memory in exchange for a small speed decrease.
|
441 |
+
|
442 |
+
Args:
|
443 |
+
slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
|
444 |
+
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
|
445 |
+
a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
|
446 |
+
`attention_head_dim` must be a multiple of `slice_size`.
|
447 |
+
"""
|
448 |
+
if slice_size == "auto":
|
449 |
+
# half the attention head size is usually a good trade-off between
|
450 |
+
# speed and memory
|
451 |
+
slice_size = self.unet.config.attention_head_dim // 2
|
452 |
+
self.unet.set_attention_slice(slice_size)
|
453 |
+
|
454 |
+
def disable_attention_slicing(self):
|
455 |
+
r"""
|
456 |
+
Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
|
457 |
+
back to computing attention in one step.
|
458 |
+
"""
|
459 |
+
# set slice_size = `None` to disable `attention slicing`
|
460 |
+
self.enable_attention_slicing(None)
|
461 |
+
|
462 |
+
@torch.no_grad()
|
463 |
+
def text2img(
|
464 |
+
self,
|
465 |
+
prompt: Union[str, List[str]],
|
466 |
+
height: int = 512,
|
467 |
+
width: int = 512,
|
468 |
+
num_inference_steps: int = 50,
|
469 |
+
guidance_scale: float = 7.5,
|
470 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
471 |
+
num_images_per_prompt: Optional[int] = 1,
|
472 |
+
eta: float = 0.0,
|
473 |
+
generator: Optional[torch.Generator] = None,
|
474 |
+
latents: Optional[torch.FloatTensor] = None,
|
475 |
+
max_embeddings_multiples: Optional[int] = 1,
|
476 |
+
output_type: Optional[str] = "pil",
|
477 |
+
return_dict: bool = True,
|
478 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
479 |
+
callback_steps: Optional[int] = 1,
|
480 |
+
**kwargs,
|
481 |
+
):
|
482 |
+
r"""
|
483 |
+
Function invoked when calling the pipeline for generation.
|
484 |
+
|
485 |
+
Args:
|
486 |
+
prompt (`str` or `List[str]`):
|
487 |
+
The prompt or prompts to guide the image generation.
|
488 |
+
height (`int`, *optional*, defaults to 512):
|
489 |
+
The height in pixels of the generated image.
|
490 |
+
width (`int`, *optional*, defaults to 512):
|
491 |
+
The width in pixels of the generated image.
|
492 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
493 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
494 |
+
expense of slower inference.
|
495 |
+
guidance_scale (`float`, *optional*, defaults to 7.5):
|
496 |
+
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
497 |
+
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
498 |
+
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
499 |
+
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
500 |
+
usually at the expense of lower image quality.
|
501 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
502 |
+
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
503 |
+
if `guidance_scale` is less than `1`).
|
504 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
505 |
+
The number of images to generate per prompt.
|
506 |
+
eta (`float`, *optional*, defaults to 0.0):
|
507 |
+
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
|
508 |
+
[`schedulers.DDIMScheduler`], will be ignored for others.
|
509 |
+
generator (`torch.Generator`, *optional*):
|
510 |
+
A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
|
511 |
+
deterministic.
|
512 |
+
latents (`torch.FloatTensor`, *optional*):
|
513 |
+
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
|
514 |
+
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
515 |
+
tensor will ge generated by sampling using the supplied random `generator`.
|
516 |
+
max_embeddings_multiples (`int`, *optional*, defaults to `1`):
|
517 |
+
The max multiple length of prompt embeddings compared to the max output length of text encoder.
|
518 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
519 |
+
The output format of the generate image. Choose between
|
520 |
+
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
521 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
522 |
+
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
|
523 |
+
plain tuple.
|
524 |
+
callback (`Callable`, *optional*):
|
525 |
+
A function that will be called every `callback_steps` steps during inference. The function will be
|
526 |
+
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
527 |
+
callback_steps (`int`, *optional*, defaults to 1):
|
528 |
+
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
529 |
+
called at every step.
|
530 |
+
|
531 |
+
Returns:
|
532 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
533 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
|
534 |
+
When returning a tuple, the first element is a list with the generated images, and the second element is a
|
535 |
+
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
|
536 |
+
(nsfw) content, according to the `safety_checker`.
|
537 |
+
"""
|
538 |
+
|
539 |
+
if isinstance(prompt, str):
|
540 |
+
batch_size = 1
|
541 |
+
elif isinstance(prompt, list):
|
542 |
+
batch_size = len(prompt)
|
543 |
+
else:
|
544 |
+
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
545 |
+
|
546 |
+
if height % 8 != 0 or width % 8 != 0:
|
547 |
+
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
|
548 |
+
|
549 |
+
if (callback_steps is None) or (
|
550 |
+
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
|
551 |
+
):
|
552 |
+
raise ValueError(
|
553 |
+
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
|
554 |
+
f" {type(callback_steps)}."
|
555 |
+
)
|
556 |
+
|
557 |
+
# get prompt text embeddings
|
558 |
+
|
559 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
560 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
561 |
+
# corresponds to doing no classifier free guidance.
|
562 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
563 |
+
# get unconditional embeddings for classifier free guidance
|
564 |
+
uncond_tokens = [""]
|
565 |
+
if do_classifier_free_guidance:
|
566 |
+
if type(prompt) is not type(negative_prompt):
|
567 |
+
raise TypeError(
|
568 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
569 |
+
f" {type(prompt)}."
|
570 |
+
)
|
571 |
+
elif isinstance(negative_prompt, str):
|
572 |
+
uncond_tokens = [negative_prompt]
|
573 |
+
elif batch_size != len(negative_prompt):
|
574 |
+
raise ValueError(
|
575 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
576 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
577 |
+
" the batch size of `prompt`."
|
578 |
+
)
|
579 |
+
else:
|
580 |
+
uncond_tokens = negative_prompt
|
581 |
+
|
582 |
+
text_embeddings = get_weighted_text_embeddings(
|
583 |
+
pipe=self,
|
584 |
+
prompt=prompt,
|
585 |
+
uncond_prompt=uncond_tokens if do_classifier_free_guidance else None,
|
586 |
+
max_embeddings_multiples=max_embeddings_multiples,
|
587 |
+
**kwargs
|
588 |
+
)
|
589 |
+
|
590 |
+
# get the initial random noise unless the user supplied it
|
591 |
+
|
592 |
+
# Unlike in other pipelines, latents need to be generated in the target device
|
593 |
+
# for 1-to-1 results reproducibility with the CompVis implementation.
|
594 |
+
# However this currently doesn't work in `mps`.
|
595 |
+
latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
|
596 |
+
latents_dtype = text_embeddings.dtype
|
597 |
+
if latents is None:
|
598 |
+
if self.device.type == "mps":
|
599 |
+
# randn does not exist on mps
|
600 |
+
latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
|
601 |
+
self.device
|
602 |
+
)
|
603 |
+
else:
|
604 |
+
latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
|
605 |
+
else:
|
606 |
+
if latents.shape != latents_shape:
|
607 |
+
raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
|
608 |
+
latents = latents.to(self.device)
|
609 |
+
|
610 |
+
# set timesteps
|
611 |
+
self.scheduler.set_timesteps(num_inference_steps)
|
612 |
+
|
613 |
+
# Some schedulers like PNDM have timesteps as arrays
|
614 |
+
# It's more optimized to move all timesteps to correct device beforehand
|
615 |
+
timesteps_tensor = self.scheduler.timesteps.to(self.device)
|
616 |
+
|
617 |
+
# scale the initial noise by the standard deviation required by the scheduler
|
618 |
+
latents = latents * self.scheduler.init_noise_sigma
|
619 |
+
|
620 |
+
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
621 |
+
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
622 |
+
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
623 |
+
# and should be between [0, 1]
|
624 |
+
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
625 |
+
extra_step_kwargs = {}
|
626 |
+
if accepts_eta:
|
627 |
+
extra_step_kwargs["eta"] = eta
|
628 |
+
|
629 |
+
for i, t in enumerate(self.progress_bar(timesteps_tensor)):
|
630 |
+
# expand the latents if we are doing classifier free guidance
|
631 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
632 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
633 |
+
|
634 |
+
# predict the noise residual
|
635 |
+
noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
|
636 |
+
|
637 |
+
# perform guidance
|
638 |
+
if do_classifier_free_guidance:
|
639 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
640 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
641 |
+
|
642 |
+
# compute the previous noisy sample x_t -> x_t-1
|
643 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
644 |
+
|
645 |
+
# call the callback, if provided
|
646 |
+
if callback is not None and i % callback_steps == 0:
|
647 |
+
callback(i, t, latents)
|
648 |
+
|
649 |
+
latents = 1 / 0.18215 * latents
|
650 |
+
image = self.vae.decode(latents).sample
|
651 |
+
|
652 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
653 |
+
|
654 |
+
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
|
655 |
+
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
656 |
+
|
657 |
+
if self.safety_checker is not None:
|
658 |
+
safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
|
659 |
+
self.device
|
660 |
+
)
|
661 |
+
image, has_nsfw_concept = self.safety_checker(
|
662 |
+
images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
|
663 |
+
)
|
664 |
+
else:
|
665 |
+
has_nsfw_concept = None
|
666 |
+
|
667 |
+
if output_type == "pil":
|
668 |
+
image = self.numpy_to_pil(image)
|
669 |
+
|
670 |
+
if not return_dict:
|
671 |
+
return (image, has_nsfw_concept)
|
672 |
+
|
673 |
+
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|
674 |
+
|
675 |
+
@torch.no_grad()
|
676 |
+
def img2img(
|
677 |
+
self,
|
678 |
+
prompt: Union[str, List[str]],
|
679 |
+
init_image: Union[torch.FloatTensor, PIL.Image.Image],
|
680 |
+
strength: float = 0.8,
|
681 |
+
num_inference_steps: Optional[int] = 50,
|
682 |
+
guidance_scale: Optional[float] = 7.5,
|
683 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
684 |
+
num_images_per_prompt: Optional[int] = 1,
|
685 |
+
eta: Optional[float] = 0.0,
|
686 |
+
generator: Optional[torch.Generator] = None,
|
687 |
+
max_embeddings_multiples: Optional[int] = 1,
|
688 |
+
output_type: Optional[str] = "pil",
|
689 |
+
return_dict: bool = True,
|
690 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
691 |
+
callback_steps: Optional[int] = 1,
|
692 |
+
**kwargs,
|
693 |
+
):
|
694 |
+
r"""
|
695 |
+
Function invoked when calling the pipeline for generation.
|
696 |
+
|
697 |
+
Args:
|
698 |
+
prompt (`str` or `List[str]`):
|
699 |
+
The prompt or prompts to guide the image generation.
|
700 |
+
init_image (`torch.FloatTensor` or `PIL.Image.Image`):
|
701 |
+
`Image`, or tensor representing an image batch, that will be used as the starting point for the
|
702 |
+
process.
|
703 |
+
strength (`float`, *optional*, defaults to 0.8):
|
704 |
+
Conceptually, indicates how much to transform the reference `init_image`. Must be between 0 and 1.
|
705 |
+
`init_image` will be used as a starting point, adding more noise to it the larger the `strength`. The
|
706 |
+
number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
|
707 |
+
noise will be maximum and the denoising process will run for the full number of iterations specified in
|
708 |
+
`num_inference_steps`. A value of 1, therefore, essentially ignores `init_image`.
|
709 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
710 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
711 |
+
expense of slower inference. This parameter will be modulated by `strength`.
|
712 |
+
guidance_scale (`float`, *optional*, defaults to 7.5):
|
713 |
+
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
714 |
+
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
715 |
+
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
716 |
+
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
717 |
+
usually at the expense of lower image quality.
|
718 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
719 |
+
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
720 |
+
if `guidance_scale` is less than `1`).
|
721 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
722 |
+
The number of images to generate per prompt.
|
723 |
+
eta (`float`, *optional*, defaults to 0.0):
|
724 |
+
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
|
725 |
+
[`schedulers.DDIMScheduler`], will be ignored for others.
|
726 |
+
generator (`torch.Generator`, *optional*):
|
727 |
+
A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
|
728 |
+
deterministic.
|
729 |
+
max_embeddings_multiples (`int`, *optional*, defaults to `1`):
|
730 |
+
The max multiple length of prompt embeddings compared to the max output length of text encoder.
|
731 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
732 |
+
The output format of the generate image. Choose between
|
733 |
+
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
734 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
735 |
+
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
|
736 |
+
plain tuple.
|
737 |
+
callback (`Callable`, *optional*):
|
738 |
+
A function that will be called every `callback_steps` steps during inference. The function will be
|
739 |
+
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
740 |
+
callback_steps (`int`, *optional*, defaults to 1):
|
741 |
+
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
742 |
+
called at every step.
|
743 |
+
|
744 |
+
Returns:
|
745 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
746 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
|
747 |
+
When returning a tuple, the first element is a list with the generated images, and the second element is a
|
748 |
+
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
|
749 |
+
(nsfw) content, according to the `safety_checker`.
|
750 |
+
"""
|
751 |
+
if isinstance(prompt, str):
|
752 |
+
batch_size = 1
|
753 |
+
elif isinstance(prompt, list):
|
754 |
+
batch_size = len(prompt)
|
755 |
+
else:
|
756 |
+
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
757 |
+
|
758 |
+
if strength < 0 or strength > 1:
|
759 |
+
raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
|
760 |
+
|
761 |
+
if (callback_steps is None) or (
|
762 |
+
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
|
763 |
+
):
|
764 |
+
raise ValueError(
|
765 |
+
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
|
766 |
+
f" {type(callback_steps)}."
|
767 |
+
)
|
768 |
+
|
769 |
+
# set timesteps
|
770 |
+
self.scheduler.set_timesteps(num_inference_steps)
|
771 |
+
|
772 |
+
if isinstance(init_image, PIL.Image.Image):
|
773 |
+
init_image = preprocess_image(init_image)
|
774 |
+
|
775 |
+
# get prompt text embeddings
|
776 |
+
|
777 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
778 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
779 |
+
# corresponds to doing no classifier free guidance.
|
780 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
781 |
+
# get unconditional embeddings for classifier free guidance
|
782 |
+
uncond_tokens = [""]
|
783 |
+
if do_classifier_free_guidance:
|
784 |
+
if type(prompt) is not type(negative_prompt):
|
785 |
+
raise TypeError(
|
786 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
787 |
+
f" {type(prompt)}."
|
788 |
+
)
|
789 |
+
elif isinstance(negative_prompt, str):
|
790 |
+
uncond_tokens = [negative_prompt]
|
791 |
+
elif batch_size != len(negative_prompt):
|
792 |
+
raise ValueError("The length of `negative_prompt` should be equal to batch_size.")
|
793 |
+
else:
|
794 |
+
uncond_tokens = negative_prompt
|
795 |
+
|
796 |
+
text_embeddings = get_weighted_text_embeddings(
|
797 |
+
pipe=self,
|
798 |
+
prompt=prompt,
|
799 |
+
uncond_prompt=uncond_tokens if do_classifier_free_guidance else None,
|
800 |
+
max_embeddings_multiples=max_embeddings_multiples,
|
801 |
+
**kwargs
|
802 |
+
)
|
803 |
+
|
804 |
+
# encode the init image into latents and scale the latents
|
805 |
+
latents_dtype = text_embeddings.dtype
|
806 |
+
init_image = init_image.to(device=self.device, dtype=latents_dtype)
|
807 |
+
init_latent_dist = self.vae.encode(init_image).latent_dist
|
808 |
+
init_latents = init_latent_dist.sample(generator=generator)
|
809 |
+
init_latents = 0.18215 * init_latents
|
810 |
+
|
811 |
+
if isinstance(prompt, str):
|
812 |
+
prompt = [prompt]
|
813 |
+
if len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] == 0:
|
814 |
+
# expand init_latents for batch_size
|
815 |
+
deprecation_message = (
|
816 |
+
f"You have passed {len(prompt)} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
|
817 |
+
" images (`init_image`). Initial images are now duplicating to match the number of text prompts. Note"
|
818 |
+
" that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
|
819 |
+
" your script to pass as many init images as text prompts to suppress this warning."
|
820 |
+
)
|
821 |
+
deprecate("len(prompt) != len(init_image)", "1.0.0", deprecation_message, standard_warn=False)
|
822 |
+
additional_image_per_prompt = len(prompt) // init_latents.shape[0]
|
823 |
+
init_latents = torch.cat([init_latents] * additional_image_per_prompt * num_images_per_prompt, dim=0)
|
824 |
+
elif len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] != 0:
|
825 |
+
raise ValueError(
|
826 |
+
f"Cannot duplicate `init_image` of batch size {init_latents.shape[0]} to {len(prompt)} text prompts."
|
827 |
+
)
|
828 |
+
else:
|
829 |
+
init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0)
|
830 |
+
|
831 |
+
# get the original timestep using init_timestep
|
832 |
+
offset = self.scheduler.config.get("steps_offset", 0)
|
833 |
+
init_timestep = int(num_inference_steps * strength) + offset
|
834 |
+
init_timestep = min(init_timestep, num_inference_steps)
|
835 |
+
|
836 |
+
timesteps = self.scheduler.timesteps[-init_timestep]
|
837 |
+
timesteps = torch.tensor([timesteps] * batch_size * num_images_per_prompt, device=self.device)
|
838 |
+
|
839 |
+
# add noise to latents using the timesteps
|
840 |
+
noise = torch.randn(init_latents.shape, generator=generator, device=self.device, dtype=latents_dtype)
|
841 |
+
init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)
|
842 |
+
|
843 |
+
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
844 |
+
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
845 |
+
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
846 |
+
# and should be between [0, 1]
|
847 |
+
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
848 |
+
extra_step_kwargs = {}
|
849 |
+
if accepts_eta:
|
850 |
+
extra_step_kwargs["eta"] = eta
|
851 |
+
|
852 |
+
latents = init_latents
|
853 |
+
|
854 |
+
t_start = max(num_inference_steps - init_timestep + offset, 0)
|
855 |
+
|
856 |
+
# Some schedulers like PNDM have timesteps as arrays
|
857 |
+
# It's more optimized to move all timesteps to correct device beforehand
|
858 |
+
timesteps = self.scheduler.timesteps[t_start:].to(self.device)
|
859 |
+
|
860 |
+
for i, t in enumerate(self.progress_bar(timesteps)):
|
861 |
+
# expand the latents if we are doing classifier free guidance
|
862 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
863 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
864 |
+
|
865 |
+
# predict the noise residual
|
866 |
+
noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
|
867 |
+
|
868 |
+
# perform guidance
|
869 |
+
if do_classifier_free_guidance:
|
870 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
871 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
872 |
+
|
873 |
+
# compute the previous noisy sample x_t -> x_t-1
|
874 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
875 |
+
|
876 |
+
# call the callback, if provided
|
877 |
+
if callback is not None and i % callback_steps == 0:
|
878 |
+
callback(i, t, latents)
|
879 |
+
|
880 |
+
latents = 1 / 0.18215 * latents
|
881 |
+
image = self.vae.decode(latents).sample
|
882 |
+
|
883 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
884 |
+
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
885 |
+
|
886 |
+
if self.safety_checker is not None:
|
887 |
+
safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
|
888 |
+
self.device
|
889 |
+
)
|
890 |
+
image, has_nsfw_concept = self.safety_checker(
|
891 |
+
images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
|
892 |
+
)
|
893 |
+
else:
|
894 |
+
has_nsfw_concept = None
|
895 |
+
|
896 |
+
if output_type == "pil":
|
897 |
+
image = self.numpy_to_pil(image)
|
898 |
+
|
899 |
+
if not return_dict:
|
900 |
+
return (image, has_nsfw_concept)
|
901 |
+
|
902 |
+
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|
903 |
+
|
904 |
+
@torch.no_grad()
|
905 |
+
def inpaint(
|
906 |
+
self,
|
907 |
+
prompt: Union[str, List[str]],
|
908 |
+
init_image: Union[torch.FloatTensor, PIL.Image.Image],
|
909 |
+
mask_image: Union[torch.FloatTensor, PIL.Image.Image],
|
910 |
+
strength: float = 0.8,
|
911 |
+
num_inference_steps: Optional[int] = 50,
|
912 |
+
guidance_scale: Optional[float] = 7.5,
|
913 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
914 |
+
num_images_per_prompt: Optional[int] = 1,
|
915 |
+
eta: Optional[float] = 0.0,
|
916 |
+
generator: Optional[torch.Generator] = None,
|
917 |
+
max_embeddings_multiples: Optional[int] = 1,
|
918 |
+
output_type: Optional[str] = "pil",
|
919 |
+
return_dict: bool = True,
|
920 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
921 |
+
callback_steps: Optional[int] = 1,
|
922 |
+
**kwargs,
|
923 |
+
):
|
924 |
+
r"""
|
925 |
+
Function invoked when calling the pipeline for generation.
|
926 |
+
|
927 |
+
Args:
|
928 |
+
prompt (`str` or `List[str]`):
|
929 |
+
The prompt or prompts to guide the image generation.
|
930 |
+
init_image (`torch.FloatTensor` or `PIL.Image.Image`):
|
931 |
+
`Image`, or tensor representing an image batch, that will be used as the starting point for the
|
932 |
+
process. This is the image whose masked region will be inpainted.
|
933 |
+
mask_image (`torch.FloatTensor` or `PIL.Image.Image`):
|
934 |
+
`Image`, or tensor representing an image batch, to mask `init_image`. White pixels in the mask will be
|
935 |
+
replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
|
936 |
+
PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should
|
937 |
+
contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.
|
938 |
+
strength (`float`, *optional*, defaults to 0.8):
|
939 |
+
Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength`
|
940 |
+
is 1, the denoising process will be run on the masked area for the full number of iterations specified
|
941 |
+
in `num_inference_steps`. `init_image` will be used as a reference for the masked area, adding more
|
942 |
+
noise to that region the larger the `strength`. If `strength` is 0, no inpainting will occur.
|
943 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
944 |
+
The reference number of denoising steps. More denoising steps usually lead to a higher quality image at
|
945 |
+
the expense of slower inference. This parameter will be modulated by `strength`, as explained above.
|
946 |
+
guidance_scale (`float`, *optional*, defaults to 7.5):
|
947 |
+
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
948 |
+
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
949 |
+
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
950 |
+
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
951 |
+
usually at the expense of lower image quality.
|
952 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
953 |
+
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
954 |
+
if `guidance_scale` is less than `1`).
|
955 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
956 |
+
The number of images to generate per prompt.
|
957 |
+
eta (`float`, *optional*, defaults to 0.0):
|
958 |
+
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
|
959 |
+
[`schedulers.DDIMScheduler`], will be ignored for others.
|
960 |
+
generator (`torch.Generator`, *optional*):
|
961 |
+
A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
|
962 |
+
deterministic.
|
963 |
+
max_embeddings_multiples (`int`, *optional*, defaults to `1`):
|
964 |
+
The max multiple length of prompt embeddings compared to the max output length of text encoder.
|
965 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
966 |
+
The output format of the generate image. Choose between
|
967 |
+
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
968 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
969 |
+
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
|
970 |
+
plain tuple.
|
971 |
+
callback (`Callable`, *optional*):
|
972 |
+
A function that will be called every `callback_steps` steps during inference. The function will be
|
973 |
+
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
974 |
+
callback_steps (`int`, *optional*, defaults to 1):
|
975 |
+
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
976 |
+
called at every step.
|
977 |
+
|
978 |
+
Returns:
|
979 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
980 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
|
981 |
+
When returning a tuple, the first element is a list with the generated images, and the second element is a
|
982 |
+
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
|
983 |
+
(nsfw) content, according to the `safety_checker`.
|
984 |
+
"""
|
985 |
+
if isinstance(prompt, str):
|
986 |
+
batch_size = 1
|
987 |
+
elif isinstance(prompt, list):
|
988 |
+
batch_size = len(prompt)
|
989 |
+
else:
|
990 |
+
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
991 |
+
|
992 |
+
if strength < 0 or strength > 1:
|
993 |
+
raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
|
994 |
+
|
995 |
+
if (callback_steps is None) or (
|
996 |
+
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
|
997 |
+
):
|
998 |
+
raise ValueError(
|
999 |
+
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
|
1000 |
+
f" {type(callback_steps)}."
|
1001 |
+
)
|
1002 |
+
|
1003 |
+
# set timesteps
|
1004 |
+
self.scheduler.set_timesteps(num_inference_steps)
|
1005 |
+
|
1006 |
+
# get prompt text embeddings
|
1007 |
+
|
1008 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
1009 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
1010 |
+
# corresponds to doing no classifier free guidance.
|
1011 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
1012 |
+
# get unconditional embeddings for classifier free guidance
|
1013 |
+
uncond_tokens = [""]
|
1014 |
+
if do_classifier_free_guidance:
|
1015 |
+
if type(prompt) is not type(negative_prompt):
|
1016 |
+
raise TypeError(
|
1017 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
1018 |
+
f" {type(prompt)}."
|
1019 |
+
)
|
1020 |
+
elif isinstance(negative_prompt, str):
|
1021 |
+
uncond_tokens = [negative_prompt]
|
1022 |
+
elif batch_size != len(negative_prompt):
|
1023 |
+
raise ValueError(
|
1024 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
1025 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
1026 |
+
" the batch size of `prompt`."
|
1027 |
+
)
|
1028 |
+
else:
|
1029 |
+
uncond_tokens = negative_prompt
|
1030 |
+
|
1031 |
+
text_embeddings = get_weighted_text_embeddings(
|
1032 |
+
pipe=self,
|
1033 |
+
prompt=prompt,
|
1034 |
+
uncond_prompt=uncond_tokens if do_classifier_free_guidance else None,
|
1035 |
+
max_embeddings_multiples=max_embeddings_multiples,
|
1036 |
+
**kwargs
|
1037 |
+
)
|
1038 |
+
|
1039 |
+
# preprocess image
|
1040 |
+
if not isinstance(init_image, torch.FloatTensor):
|
1041 |
+
init_image = preprocess_image(init_image)
|
1042 |
+
|
1043 |
+
# encode the init image into latents and scale the latents
|
1044 |
+
latents_dtype = text_embeddings.dtype
|
1045 |
+
init_image = init_image.to(device=self.device, dtype=latents_dtype)
|
1046 |
+
init_latent_dist = self.vae.encode(init_image).latent_dist
|
1047 |
+
init_latents = init_latent_dist.sample(generator=generator)
|
1048 |
+
init_latents = 0.18215 * init_latents
|
1049 |
+
|
1050 |
+
# Expand init_latents for batch_size and num_images_per_prompt
|
1051 |
+
init_latents = torch.cat([init_latents] * batch_size * num_images_per_prompt, dim=0)
|
1052 |
+
init_latents_orig = init_latents
|
1053 |
+
|
1054 |
+
# preprocess mask
|
1055 |
+
if not isinstance(mask_image, torch.FloatTensor):
|
1056 |
+
mask_image = preprocess_mask(mask_image)
|
1057 |
+
mask_image = mask_image.to(device=self.device, dtype=latents_dtype)
|
1058 |
+
mask = torch.cat([mask_image] * batch_size * num_images_per_prompt)
|
1059 |
+
|
1060 |
+
# check sizes
|
1061 |
+
if not mask.shape == init_latents.shape:
|
1062 |
+
raise ValueError("The mask and init_image should be the same size!")
|
1063 |
+
|
1064 |
+
# get the original timestep using init_timestep
|
1065 |
+
offset = self.scheduler.config.get("steps_offset", 0)
|
1066 |
+
init_timestep = int(num_inference_steps * strength) + offset
|
1067 |
+
init_timestep = min(init_timestep, num_inference_steps)
|
1068 |
+
|
1069 |
+
timesteps = self.scheduler.timesteps[-init_timestep]
|
1070 |
+
timesteps = torch.tensor([timesteps] * batch_size * num_images_per_prompt, device=self.device)
|
1071 |
+
|
1072 |
+
# add noise to latents using the timesteps
|
1073 |
+
noise = torch.randn(init_latents.shape, generator=generator, device=self.device, dtype=latents_dtype)
|
1074 |
+
init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)
|
1075 |
+
|
1076 |
+
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
1077 |
+
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
1078 |
+
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
1079 |
+
# and should be between [0, 1]
|
1080 |
+
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
1081 |
+
extra_step_kwargs = {}
|
1082 |
+
if accepts_eta:
|
1083 |
+
extra_step_kwargs["eta"] = eta
|
1084 |
+
|
1085 |
+
latents = init_latents
|
1086 |
+
|
1087 |
+
t_start = max(num_inference_steps - init_timestep + offset, 0)
|
1088 |
+
|
1089 |
+
# Some schedulers like PNDM have timesteps as arrays
|
1090 |
+
# It's more optimized to move all timesteps to correct device beforehand
|
1091 |
+
timesteps = self.scheduler.timesteps[t_start:].to(self.device)
|
1092 |
+
|
1093 |
+
for i, t in enumerate(self.progress_bar(timesteps)):
|
1094 |
+
# expand the latents if we are doing classifier free guidance
|
1095 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
1096 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
1097 |
+
|
1098 |
+
# predict the noise residual
|
1099 |
+
noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
|
1100 |
+
|
1101 |
+
# perform guidance
|
1102 |
+
if do_classifier_free_guidance:
|
1103 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
1104 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
1105 |
+
|
1106 |
+
# compute the previous noisy sample x_t -> x_t-1
|
1107 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
1108 |
+
# masking
|
1109 |
+
init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, torch.tensor([t]))
|
1110 |
+
|
1111 |
+
latents = (init_latents_proper * mask) + (latents * (1 - mask))
|
1112 |
+
|
1113 |
+
# call the callback, if provided
|
1114 |
+
if callback is not None and i % callback_steps == 0:
|
1115 |
+
callback(i, t, latents)
|
1116 |
+
|
1117 |
+
latents = 1 / 0.18215 * latents
|
1118 |
+
image = self.vae.decode(latents).sample
|
1119 |
+
|
1120 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
1121 |
+
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
1122 |
+
|
1123 |
+
if self.safety_checker is not None:
|
1124 |
+
safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
|
1125 |
+
self.device
|
1126 |
+
)
|
1127 |
+
image, has_nsfw_concept = self.safety_checker(images=image, clip_input=safety_checker_input.pixel_values)
|
1128 |
+
else:
|
1129 |
+
has_nsfw_concept = None
|
1130 |
+
|
1131 |
+
if output_type == "pil":
|
1132 |
+
image = self.numpy_to_pil(image)
|
1133 |
+
|
1134 |
+
if not return_dict:
|
1135 |
+
return (image, has_nsfw_concept)
|
1136 |
+
|
1137 |
+
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|