Spaces:
Runtime error
Runtime error
Shanghua Gao
commited on
Commit
·
78acc58
1
Parent(s):
5fb4baa
udpate
Browse files- README.md +0 -1
- annotator/util.py +23 -1
- app.py +1 -1
- editany_demo.py +157 -85
- editany_lora.py +78 -44
- editany_nogradio.py +20 -0
- editany_test.py +1 -1
- environment.yaml +38 -0
- requirements.txt +1 -1
- utils/stable_diffusion_controlnet_inpaint.py +26 -18
- utils/stable_diffusion_reference.py +295 -326
README.md
CHANGED
@@ -8,7 +8,6 @@ sdk_version: 3.35.2
|
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
---
|
11 |
-
|
12 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
13 |
|
14 |
# Edit Anything by Segment-Anything
|
|
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
---
|
|
|
11 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
12 |
|
13 |
# Edit Anything by Segment-Anything
|
annotator/util.py
CHANGED
@@ -1,7 +1,7 @@
|
|
1 |
import numpy as np
|
2 |
import cv2
|
3 |
import os
|
4 |
-
|
5 |
|
6 |
annotator_ckpts_path = os.path.join(os.path.dirname(__file__), 'ckpts')
|
7 |
|
@@ -71,3 +71,25 @@ def get_bounding_box(mask):
|
|
71 |
|
72 |
# Return as [xmin, ymin, xmax, ymax]
|
73 |
return [rmin, cmin, rmax, cmax]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import numpy as np
|
2 |
import cv2
|
3 |
import os
|
4 |
+
import pickle
|
5 |
|
6 |
annotator_ckpts_path = os.path.join(os.path.dirname(__file__), 'ckpts')
|
7 |
|
|
|
71 |
|
72 |
# Return as [xmin, ymin, xmax, ymax]
|
73 |
return [rmin, cmin, rmax, cmax]
|
74 |
+
|
75 |
+
|
76 |
+
|
77 |
+
def save_input_to_file(func):
|
78 |
+
def wrapper(self, *args, **kwargs):
|
79 |
+
# 创建不包含 self 的输入副本
|
80 |
+
input_data = {
|
81 |
+
'args': args,
|
82 |
+
'kwargs': kwargs
|
83 |
+
}
|
84 |
+
|
85 |
+
# 执行原始函数
|
86 |
+
result = func(self, *args, **kwargs)
|
87 |
+
|
88 |
+
# 将输入数据保存到文件
|
89 |
+
with open('input_data.pkl', 'wb') as f:
|
90 |
+
pickle.dump(input_data, f)
|
91 |
+
|
92 |
+
# 返回结果
|
93 |
+
return result
|
94 |
+
|
95 |
+
return wrapper
|
app.py
CHANGED
@@ -68,4 +68,4 @@ with gr.Blocks() as demo:
|
|
68 |
with gr.Tabs():
|
69 |
gr.Markdown(SHARED_UI_WARNING)
|
70 |
|
71 |
-
demo.queue(api_open=False).launch(server_name='0.0.0.0', share=False)
|
|
|
68 |
with gr.Tabs():
|
69 |
gr.Markdown(SHARED_UI_WARNING)
|
70 |
|
71 |
+
demo.queue(api_open=False).launch(server_name='0.0.0.0', share=False)
|
editany_demo.py
CHANGED
@@ -1,6 +1,10 @@
|
|
1 |
# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2
|
2 |
import gradio as gr
|
3 |
|
|
|
|
|
|
|
|
|
4 |
|
5 |
def create_demo_template(
|
6 |
process,
|
@@ -22,7 +26,7 @@ def create_demo_template(
|
|
22 |
ref_click_mask = gr.State(None)
|
23 |
with gr.Row():
|
24 |
gr.Markdown(INFO)
|
25 |
-
with gr.Row(
|
26 |
with gr.Column():
|
27 |
with gr.Tab("Click🖱"):
|
28 |
source_image_click = gr.Image(
|
@@ -40,12 +44,13 @@ def create_demo_template(
|
|
40 |
interactive=True,
|
41 |
show_label=False,
|
42 |
)
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
|
|
49 |
with gr.Row():
|
50 |
run_button_click = gr.Button(
|
51 |
label="Run EditAnying", interactive=True
|
@@ -56,63 +61,75 @@ def create_demo_template(
|
|
56 |
label="Image: Upload an image and cover the region you want to edit with sketch",
|
57 |
type="numpy",
|
58 |
tool="sketch",
|
|
|
59 |
)
|
60 |
run_button = gr.Button(
|
61 |
label="Run EditAnying", interactive=True)
|
62 |
-
with gr.
|
63 |
-
|
64 |
-
|
|
|
|
|
65 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
control_scale = gr.Slider(
|
67 |
-
label="Mask
|
68 |
-
info="Large value -> strict alignment with SAM mask",
|
69 |
minimum=0,
|
70 |
maximum=1,
|
71 |
value=0.5,
|
72 |
step=0.1,
|
73 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
with gr.Column():
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
)
|
93 |
-
seed = gr.Slider(
|
94 |
-
label="Seed",
|
95 |
-
minimum=-1,
|
96 |
-
maximum=2147483647,
|
97 |
-
step=1,
|
98 |
-
randomize=True,
|
99 |
-
)
|
100 |
with gr.Row():
|
101 |
enable_tile = gr.Checkbox(
|
102 |
-
label="
|
103 |
info="Slow inference",
|
104 |
value=True,
|
105 |
)
|
106 |
refine_alignment_ratio = gr.Slider(
|
107 |
-
label="
|
108 |
-
info="Large value -> strict alignment with input image. Small value -> strong global consistency",
|
109 |
minimum=0.0,
|
110 |
maximum=1.0,
|
111 |
value=0.95,
|
112 |
step=0.05,
|
113 |
)
|
114 |
|
115 |
-
with gr.Accordion("
|
116 |
# ref_image = gr.Image(
|
117 |
# source='upload', label="Upload a reference image", type="pil", value=None)
|
118 |
ref_image = gr.Image(
|
@@ -120,8 +137,9 @@ def create_demo_template(
|
|
120 |
label="Upload a reference image and cover the region you want to use with sketch",
|
121 |
type="pil",
|
122 |
tool="sketch",
|
|
|
123 |
)
|
124 |
-
with gr.
|
125 |
ref_auto_prompt = gr.Checkbox(
|
126 |
label="Ref. Auto Prompt", value=True
|
127 |
)
|
@@ -148,45 +166,25 @@ def create_demo_template(
|
|
148 |
with gr.Row():
|
149 |
reference_attn = gr.Checkbox(
|
150 |
label="reference_attn", value=True)
|
151 |
-
|
152 |
-
label="
|
153 |
-
minimum=0,
|
154 |
-
maximum=1.0,
|
155 |
-
value=0.8,
|
156 |
-
step=0.01,
|
157 |
)
|
158 |
with gr.Row():
|
159 |
-
|
160 |
-
label="
|
|
|
|
|
|
|
|
|
161 |
)
|
162 |
-
|
163 |
-
label="
|
164 |
minimum=0,
|
165 |
maximum=1.0,
|
166 |
-
value=0.
|
167 |
-
step=0.
|
168 |
)
|
169 |
-
|
170 |
-
label="Style fidelity",
|
171 |
-
minimum=0,
|
172 |
-
maximum=1.0,
|
173 |
-
value=0.5,
|
174 |
-
step=0.01,
|
175 |
-
)
|
176 |
-
ref_sam_scale = gr.Slider(
|
177 |
-
label="SAM Control Scale",
|
178 |
-
minimum=0,
|
179 |
-
maximum=1.0,
|
180 |
-
value=0.3,
|
181 |
-
step=0.1,
|
182 |
-
)
|
183 |
-
ref_inpaint_scale = gr.Slider(
|
184 |
-
label="Inpaint Control Scale",
|
185 |
-
minimum=0,
|
186 |
-
maximum=1.0,
|
187 |
-
value=0.2,
|
188 |
-
step=0.1,
|
189 |
-
)
|
190 |
with gr.Row():
|
191 |
ref_textinv = gr.Checkbox(
|
192 |
label="Use textual inversion token", value=False
|
@@ -196,8 +194,37 @@ def create_demo_template(
|
|
196 |
info="Text in the inversion token path",
|
197 |
value=None,
|
198 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
199 |
|
200 |
-
with gr.Accordion("Advanced
|
201 |
mask_image = gr.Image(
|
202 |
source="upload",
|
203 |
label="Upload a predefined mask of edit region: Switch to Brush mode when using this!",
|
@@ -244,19 +271,16 @@ def create_demo_template(
|
|
244 |
)
|
245 |
with gr.Column():
|
246 |
result_gallery_refine = gr.Gallery(
|
247 |
-
label="Output High quality", show_label=True, elem_id="gallery"
|
248 |
-
).style(grid=2, preview=False)
|
249 |
result_gallery_init = gr.Gallery(
|
250 |
-
label="Output Low quality", show_label=True, elem_id="gallery"
|
251 |
-
).style(grid=2, height="auto")
|
252 |
result_gallery_ref = gr.Gallery(
|
253 |
-
label="Output Ref", show_label=False, elem_id="gallery"
|
254 |
-
|
255 |
-
result_text = gr.Text(label="BLIP2+Human Prompt Text")
|
256 |
|
257 |
ips = [
|
258 |
source_image_brush,
|
259 |
-
enable_all_generate
|
260 |
mask_image,
|
261 |
control_scale,
|
262 |
enable_auto_prompt,
|
@@ -288,6 +312,7 @@ def create_demo_template(
|
|
288 |
ref_auto_prompt,
|
289 |
ref_textinv,
|
290 |
ref_textinv_path,
|
|
|
291 |
]
|
292 |
run_button.click(
|
293 |
fn=process,
|
@@ -299,10 +324,56 @@ def create_demo_template(
|
|
299 |
result_text,
|
300 |
],
|
301 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
302 |
|
303 |
ip_click = [
|
304 |
origin_image,
|
305 |
-
enable_all_generate
|
306 |
click_mask,
|
307 |
control_scale,
|
308 |
enable_auto_prompt,
|
@@ -334,6 +405,7 @@ def create_demo_template(
|
|
334 |
ref_auto_prompt,
|
335 |
ref_textinv,
|
336 |
ref_textinv_path,
|
|
|
337 |
]
|
338 |
|
339 |
run_button_click.click(
|
|
|
1 |
# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2
|
2 |
import gradio as gr
|
3 |
|
4 |
+
import numpy as np
|
5 |
+
import cv2
|
6 |
+
from cv2 import imencode
|
7 |
+
import base64
|
8 |
|
9 |
def create_demo_template(
|
10 |
process,
|
|
|
26 |
ref_click_mask = gr.State(None)
|
27 |
with gr.Row():
|
28 |
gr.Markdown(INFO)
|
29 |
+
with gr.Row(equal_height=False):
|
30 |
with gr.Column():
|
31 |
with gr.Tab("Click🖱"):
|
32 |
source_image_click = gr.Image(
|
|
|
44 |
interactive=True,
|
45 |
show_label=False,
|
46 |
)
|
47 |
+
with gr.Row():
|
48 |
+
clear_button_click = gr.Button(
|
49 |
+
value="Clear Points", interactive=True
|
50 |
+
)
|
51 |
+
clear_button_image = gr.Button(
|
52 |
+
value="Reset Image", interactive=True
|
53 |
+
)
|
54 |
with gr.Row():
|
55 |
run_button_click = gr.Button(
|
56 |
label="Run EditAnying", interactive=True
|
|
|
61 |
label="Image: Upload an image and cover the region you want to edit with sketch",
|
62 |
type="numpy",
|
63 |
tool="sketch",
|
64 |
+
brush_color="#00FFBF"
|
65 |
)
|
66 |
run_button = gr.Button(
|
67 |
label="Run EditAnying", interactive=True)
|
68 |
+
with gr.Tab("All region"):
|
69 |
+
source_image_clean = gr.Image(
|
70 |
+
source="upload",
|
71 |
+
label="Image: Upload an image",
|
72 |
+
type="numpy",
|
73 |
)
|
74 |
+
run_button_allregion = gr.Button(
|
75 |
+
label="Run EditAnying", interactive=True)
|
76 |
+
with gr.Row():
|
77 |
+
# enable_all_generate = gr.Checkbox(
|
78 |
+
# label="All Region Generation", value=False
|
79 |
+
# )
|
80 |
control_scale = gr.Slider(
|
81 |
+
label="SAM Mask Alignment Strength",
|
82 |
+
# info="Large value -> strict alignment with SAM mask",
|
83 |
minimum=0,
|
84 |
maximum=1,
|
85 |
value=0.5,
|
86 |
step=0.1,
|
87 |
)
|
88 |
+
with gr.Row():
|
89 |
+
num_samples = gr.Slider(
|
90 |
+
label="Images", minimum=1, maximum=12, value=2, step=1
|
91 |
+
)
|
92 |
+
seed = gr.Slider(
|
93 |
+
label="Seed",
|
94 |
+
minimum=-1,
|
95 |
+
maximum=2147483647,
|
96 |
+
step=1,
|
97 |
+
randomize=True,
|
98 |
+
)
|
99 |
with gr.Column():
|
100 |
+
with gr.Row():
|
101 |
+
enable_auto_prompt = gr.Checkbox(
|
102 |
+
label="Prompt Auto Generation (Enable this may makes your prompt not working)",
|
103 |
+
# info="",
|
104 |
+
value=enable_auto_prompt_default,
|
105 |
+
)
|
106 |
+
with gr.Row():
|
107 |
+
a_prompt = gr.Textbox(
|
108 |
+
label="Positive Prompt",
|
109 |
+
info="Text in the expected things of edited region",
|
110 |
+
value="best quality, extremely detailed,",
|
111 |
+
)
|
112 |
+
n_prompt = gr.Textbox(
|
113 |
+
label="Negative Prompt",
|
114 |
+
value="longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, NSFW",
|
115 |
+
)
|
116 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
117 |
with gr.Row():
|
118 |
enable_tile = gr.Checkbox(
|
119 |
+
label="High-resolution Refinement",
|
120 |
info="Slow inference",
|
121 |
value=True,
|
122 |
)
|
123 |
refine_alignment_ratio = gr.Slider(
|
124 |
+
label="Similarity with Initial Results",
|
125 |
+
# info="Large value -> strict alignment with input image. Small value -> strong global consistency",
|
126 |
minimum=0.0,
|
127 |
maximum=1.0,
|
128 |
value=0.95,
|
129 |
step=0.05,
|
130 |
)
|
131 |
|
132 |
+
with gr.Accordion("Cross-image Drag Options", open=False):
|
133 |
# ref_image = gr.Image(
|
134 |
# source='upload', label="Upload a reference image", type="pil", value=None)
|
135 |
ref_image = gr.Image(
|
|
|
137 |
label="Upload a reference image and cover the region you want to use with sketch",
|
138 |
type="pil",
|
139 |
tool="sketch",
|
140 |
+
brush_color="#00FFBF",
|
141 |
)
|
142 |
+
with gr.Row():
|
143 |
ref_auto_prompt = gr.Checkbox(
|
144 |
label="Ref. Auto Prompt", value=True
|
145 |
)
|
|
|
166 |
with gr.Row():
|
167 |
reference_attn = gr.Checkbox(
|
168 |
label="reference_attn", value=True)
|
169 |
+
reference_adain = gr.Checkbox(
|
170 |
+
label="reference_adain", value=True
|
|
|
|
|
|
|
|
|
171 |
)
|
172 |
with gr.Row():
|
173 |
+
ref_sam_scale = gr.Slider(
|
174 |
+
label="Pos Control Scale",
|
175 |
+
minimum=0,
|
176 |
+
maximum=1.0,
|
177 |
+
value=0.3,
|
178 |
+
step=0.1,
|
179 |
)
|
180 |
+
ref_inpaint_scale = gr.Slider(
|
181 |
+
label="Content Control Scale",
|
182 |
minimum=0,
|
183 |
maximum=1.0,
|
184 |
+
value=0.2,
|
185 |
+
step=0.1,
|
186 |
)
|
187 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
with gr.Row():
|
189 |
ref_textinv = gr.Checkbox(
|
190 |
label="Use textual inversion token", value=False
|
|
|
194 |
info="Text in the inversion token path",
|
195 |
value=None,
|
196 |
)
|
197 |
+
with gr.Accordion("Advanced options", open=False):
|
198 |
+
style_fidelity = gr.Slider(
|
199 |
+
label="Style fidelity",
|
200 |
+
minimum=0,
|
201 |
+
maximum=1.,
|
202 |
+
value=0.,
|
203 |
+
step=0.1,
|
204 |
+
)
|
205 |
+
attention_auto_machine_weight = gr.Slider(
|
206 |
+
label="Attention Reference Weight",
|
207 |
+
minimum=0,
|
208 |
+
maximum=1.0,
|
209 |
+
value=1.0,
|
210 |
+
step=0.01,
|
211 |
+
)
|
212 |
+
gn_auto_machine_weight = gr.Slider(
|
213 |
+
label="GroupNorm Reference Weight",
|
214 |
+
minimum=0,
|
215 |
+
maximum=1.0,
|
216 |
+
value=1.0,
|
217 |
+
step=0.01,
|
218 |
+
)
|
219 |
+
ref_scale = gr.Slider(
|
220 |
+
label="Frequency Reference Guidance Scale",
|
221 |
+
minimum=0,
|
222 |
+
maximum=1.0,
|
223 |
+
value=0.0,
|
224 |
+
step=0.1,
|
225 |
+
)
|
226 |
|
227 |
+
with gr.Accordion("Advanced Options", open=False):
|
228 |
mask_image = gr.Image(
|
229 |
source="upload",
|
230 |
label="Upload a predefined mask of edit region: Switch to Brush mode when using this!",
|
|
|
271 |
)
|
272 |
with gr.Column():
|
273 |
result_gallery_refine = gr.Gallery(
|
274 |
+
label="Output High quality", show_label=True, elem_id="gallery", preview=False)
|
|
|
275 |
result_gallery_init = gr.Gallery(
|
276 |
+
label="Output Low quality", show_label=True, elem_id="gallery", height="auto")
|
|
|
277 |
result_gallery_ref = gr.Gallery(
|
278 |
+
label="Output Ref", show_label=False, elem_id="gallery", height="auto")
|
279 |
+
result_text = gr.Text(label="ALL Prompt Text")
|
|
|
280 |
|
281 |
ips = [
|
282 |
source_image_brush,
|
283 |
+
gr.State(False), # enable_all_generate
|
284 |
mask_image,
|
285 |
control_scale,
|
286 |
enable_auto_prompt,
|
|
|
312 |
ref_auto_prompt,
|
313 |
ref_textinv,
|
314 |
ref_textinv_path,
|
315 |
+
ref_scale,
|
316 |
]
|
317 |
run_button.click(
|
318 |
fn=process,
|
|
|
324 |
result_text,
|
325 |
],
|
326 |
)
|
327 |
+
ips_allregion = [
|
328 |
+
source_image_clean,
|
329 |
+
gr.State(True), # enable_all_generate
|
330 |
+
mask_image,
|
331 |
+
control_scale,
|
332 |
+
enable_auto_prompt,
|
333 |
+
a_prompt,
|
334 |
+
n_prompt,
|
335 |
+
num_samples,
|
336 |
+
image_resolution,
|
337 |
+
detect_resolution,
|
338 |
+
ddim_steps,
|
339 |
+
guess_mode,
|
340 |
+
scale,
|
341 |
+
seed,
|
342 |
+
eta,
|
343 |
+
enable_tile,
|
344 |
+
refine_alignment_ratio,
|
345 |
+
refine_image_resolution,
|
346 |
+
alpha_weight,
|
347 |
+
use_scale_map,
|
348 |
+
condition_model,
|
349 |
+
ref_image,
|
350 |
+
attention_auto_machine_weight,
|
351 |
+
gn_auto_machine_weight,
|
352 |
+
style_fidelity,
|
353 |
+
reference_attn,
|
354 |
+
reference_adain,
|
355 |
+
ref_prompt,
|
356 |
+
ref_sam_scale,
|
357 |
+
ref_inpaint_scale,
|
358 |
+
ref_auto_prompt,
|
359 |
+
ref_textinv,
|
360 |
+
ref_textinv_path,
|
361 |
+
ref_scale,
|
362 |
+
]
|
363 |
+
run_button_allregion.click(
|
364 |
+
fn=process,
|
365 |
+
inputs=ips_allregion,
|
366 |
+
outputs=[
|
367 |
+
result_gallery_refine,
|
368 |
+
result_gallery_init,
|
369 |
+
result_gallery_ref,
|
370 |
+
result_text,
|
371 |
+
],
|
372 |
+
)
|
373 |
|
374 |
ip_click = [
|
375 |
origin_image,
|
376 |
+
gr.State(False), # enable_all_generate
|
377 |
click_mask,
|
378 |
control_scale,
|
379 |
enable_auto_prompt,
|
|
|
405 |
ref_auto_prompt,
|
406 |
ref_textinv,
|
407 |
ref_textinv_path,
|
408 |
+
ref_scale,
|
409 |
]
|
410 |
|
411 |
run_button_click.click(
|
editany_lora.py
CHANGED
@@ -14,7 +14,7 @@ import random
|
|
14 |
import os
|
15 |
import requests
|
16 |
from io import BytesIO
|
17 |
-
from annotator.util import resize_image, HWC3, resize_points, get_bounding_box
|
18 |
|
19 |
import torch
|
20 |
from safetensors.torch import load_file
|
@@ -28,8 +28,7 @@ from utils.stable_diffusion_controlnet_inpaint import StableDiffusionControlNetI
|
|
28 |
# need the latest transformers
|
29 |
# pip install git+https://github.com/huggingface/transformers.git
|
30 |
from transformers import AutoProcessor, Blip2ForConditionalGeneration
|
31 |
-
from diffusers import ControlNetModel
|
32 |
-
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
|
33 |
import PIL.Image
|
34 |
|
35 |
# Segment-Anything init.
|
@@ -119,16 +118,55 @@ def get_pipeline_embeds(pipeline, prompt, negative_prompt, device):
|
|
119 |
"""
|
120 |
max_length = pipeline.tokenizer.model_max_length
|
121 |
|
122 |
-
# simple way to determine length of tokens
|
123 |
-
count_prompt = len(re.split(r",
|
124 |
-
count_negative_prompt = len(re.split(r",
|
125 |
-
|
126 |
-
# create the tensor based on which prompt is longer
|
127 |
-
if count_prompt >= count_negative_prompt:
|
128 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
prompt, return_tensors="pt", truncation=False
|
130 |
).input_ids.to(device)
|
131 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
132 |
negative_ids = pipeline.tokenizer(
|
133 |
negative_prompt,
|
134 |
truncation=False,
|
@@ -137,23 +175,21 @@ def get_pipeline_embeds(pipeline, prompt, negative_prompt, device):
|
|
137 |
return_tensors="pt",
|
138 |
).input_ids.to(device)
|
139 |
else:
|
140 |
-
negative_ids = pipeline.tokenizer(
|
141 |
-
negative_prompt, return_tensors="pt", truncation=False
|
142 |
-
).input_ids.to(device)
|
143 |
-
shape_max_length = negative_ids.shape[-1]
|
144 |
input_ids = pipeline.tokenizer(
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
|
152 |
concat_embeds = []
|
153 |
neg_embeds = []
|
154 |
for i in range(0, shape_max_length, max_length):
|
155 |
-
concat_embeds.append(pipeline.text_encoder(
|
156 |
-
|
|
|
|
|
157 |
|
158 |
return torch.cat(concat_embeds, dim=1), torch.cat(neg_embeds, dim=1)
|
159 |
|
@@ -178,10 +214,12 @@ def load_lora_weights(pipeline, checkpoint_path, multiplier, device, dtype):
|
|
178 |
for layer, elems in updates.items():
|
179 |
|
180 |
if "text" in layer:
|
181 |
-
layer_infos = layer.split(
|
|
|
182 |
curr_layer = pipeline.text_encoder
|
183 |
else:
|
184 |
-
layer_infos = layer.split(
|
|
|
185 |
curr_layer = pipeline.unet
|
186 |
|
187 |
# find the target layer
|
@@ -244,7 +282,8 @@ def load_lora_weights(pipeline, checkpoint_path, multiplier, device, dtype):
|
|
244 |
)
|
245 |
curr_layer = pipeline.text_encoder
|
246 |
else:
|
247 |
-
layer_infos = layer.split(
|
|
|
248 |
curr_layer = pipeline.unet
|
249 |
|
250 |
# find the target layer
|
@@ -489,7 +528,7 @@ class EditAnythingLoraModel:
|
|
489 |
self.mask_predictor.set_image(image)
|
490 |
# Separate the points and labels
|
491 |
points, labels = zip(*[(point[:2], point[2])
|
492 |
-
|
493 |
|
494 |
# Convert the points and labels to numpy arrays
|
495 |
input_point = np.array(points)
|
@@ -534,7 +573,8 @@ class EditAnythingLoraModel:
|
|
534 |
mask_click_np = np.transpose(mask_click_np, (1, 2, 0)) * 255.0
|
535 |
|
536 |
mask_image = HWC3(mask_click_np.astype(np.uint8))
|
537 |
-
mask_image = cv2.resize(
|
|
|
538 |
# mask_image = Image.fromarray(mask_image_tmp)
|
539 |
|
540 |
# Draw circles for all clicked points
|
@@ -567,6 +607,7 @@ class EditAnythingLoraModel:
|
|
567 |
)
|
568 |
|
569 |
@torch.inference_mode()
|
|
|
570 |
def process(
|
571 |
self,
|
572 |
source_image,
|
@@ -602,6 +643,7 @@ class EditAnythingLoraModel:
|
|
602 |
ref_auto_prompt=False,
|
603 |
ref_textinv=True,
|
604 |
ref_textinv_path=None,
|
|
|
605 |
):
|
606 |
|
607 |
if condition_model is None or condition_model == "EditAnything":
|
@@ -624,14 +666,9 @@ class EditAnythingLoraModel:
|
|
624 |
)
|
625 |
self.defalut_enable_all_generate = enable_all_generate
|
626 |
if enable_all_generate:
|
627 |
-
print(
|
628 |
-
"source_image",
|
629 |
-
source_image["mask"].shape,
|
630 |
-
input_image.shape,
|
631 |
-
)
|
632 |
mask_image = (
|
633 |
np.ones((input_image.shape[0],
|
634 |
-
|
635 |
)
|
636 |
else:
|
637 |
mask_image = source_image["mask"]
|
@@ -699,11 +736,13 @@ class EditAnythingLoraModel:
|
|
699 |
except:
|
700 |
print("No textinvert embeddings found.")
|
701 |
ref_data_path = "./utils/tmp/textinv/img"
|
702 |
-
if not os.path.exists(ref_data_path):
|
703 |
os.makedirs(ref_data_path)
|
704 |
-
cropped_ref_image.save(
|
|
|
705 |
print("Ref image region is save to:", ref_data_path)
|
706 |
-
print(
|
|
|
707 |
|
708 |
else:
|
709 |
ref_mask = None
|
@@ -735,7 +774,7 @@ class EditAnythingLoraModel:
|
|
735 |
)
|
736 |
|
737 |
control = torch.from_numpy(detected_map.copy()).float().cuda()
|
738 |
-
control =
|
739 |
control = einops.rearrange(control, "b h w c -> b c h w").clone()
|
740 |
|
741 |
mask_imag_ori = HWC3(mask_image.astype(np.uint8))
|
@@ -753,14 +792,8 @@ class EditAnythingLoraModel:
|
|
753 |
prompt_embeds, negative_prompt_embeds = get_pipeline_embeds(
|
754 |
self.pipe, postive_prompt, negative_prompt, "cuda"
|
755 |
)
|
756 |
-
prompt_embeds = torch.cat([prompt_embeds] * num_samples, dim=0)
|
757 |
-
negative_prompt_embeds = torch.cat(
|
758 |
-
[negative_prompt_embeds] * num_samples, dim=0
|
759 |
-
)
|
760 |
|
761 |
if enable_all_generate and self.extra_inpaint:
|
762 |
-
self.pipe.safety_checker = lambda images, clip_input: (
|
763 |
-
images, False)
|
764 |
if ref_image is not None:
|
765 |
print("Not support yet.")
|
766 |
return
|
@@ -845,6 +878,7 @@ class EditAnythingLoraModel:
|
|
845 |
reference_adain=reference_adain,
|
846 |
ref_controlnet_conditioning_scale=ref_multi_condition_scale,
|
847 |
guess_mode=guess_mode,
|
|
|
848 |
).images
|
849 |
results = [x_samples[i] for i in range(num_samples)]
|
850 |
|
|
|
14 |
import os
|
15 |
import requests
|
16 |
from io import BytesIO
|
17 |
+
from annotator.util import resize_image, HWC3, resize_points, get_bounding_box, save_input_to_file
|
18 |
|
19 |
import torch
|
20 |
from safetensors.torch import load_file
|
|
|
28 |
# need the latest transformers
|
29 |
# pip install git+https://github.com/huggingface/transformers.git
|
30 |
from transformers import AutoProcessor, Blip2ForConditionalGeneration
|
31 |
+
from diffusers import ControlNetModel
|
|
|
32 |
import PIL.Image
|
33 |
|
34 |
# Segment-Anything init.
|
|
|
118 |
"""
|
119 |
max_length = pipeline.tokenizer.model_max_length
|
120 |
|
121 |
+
# # simple way to determine length of tokens
|
122 |
+
# count_prompt = len(re.split(r",", prompt))
|
123 |
+
# count_negative_prompt = len(re.split(r",", negative_prompt))
|
124 |
+
|
125 |
+
# # create the tensor based on which prompt is longer
|
126 |
+
# if count_prompt >= count_negative_prompt:
|
127 |
+
# input_ids = pipeline.tokenizer(
|
128 |
+
# prompt, return_tensors="pt", truncation=False
|
129 |
+
# ).input_ids.to(device)
|
130 |
+
# shape_max_length = input_ids.shape[-1]
|
131 |
+
# negative_ids = pipeline.tokenizer(
|
132 |
+
# negative_prompt,
|
133 |
+
# truncation=False,
|
134 |
+
# padding="max_length",
|
135 |
+
# max_length=shape_max_length,
|
136 |
+
# return_tensors="pt",
|
137 |
+
# ).input_ids.to(device)
|
138 |
+
# else:
|
139 |
+
# negative_ids = pipeline.tokenizer(
|
140 |
+
# negative_prompt, return_tensors="pt", truncation=False
|
141 |
+
# ).input_ids.to(device)
|
142 |
+
# shape_max_length = negative_ids.shape[-1]
|
143 |
+
# input_ids = pipeline.tokenizer(
|
144 |
+
# prompt,
|
145 |
+
# return_tensors="pt",
|
146 |
+
# truncation=False,
|
147 |
+
# padding="max_length",
|
148 |
+
# max_length=shape_max_length,
|
149 |
+
# ).input_ids.to(device)
|
150 |
+
|
151 |
+
# concat_embeds = []
|
152 |
+
# neg_embeds = []
|
153 |
+
# for i in range(0, shape_max_length, max_length):
|
154 |
+
# concat_embeds.append(pipeline.text_encoder(
|
155 |
+
# input_ids[:, i: i + max_length])[0])
|
156 |
+
# neg_embeds.append(pipeline.text_encoder(
|
157 |
+
# negative_ids[:, i: i + max_length])[0])
|
158 |
+
|
159 |
+
input_ids = pipeline.tokenizer(
|
160 |
prompt, return_tensors="pt", truncation=False
|
161 |
).input_ids.to(device)
|
162 |
+
|
163 |
+
negative_ids = pipeline.tokenizer(
|
164 |
+
negative_prompt, return_tensors="pt", truncation=False
|
165 |
+
).input_ids.to(device)
|
166 |
+
|
167 |
+
shape_max_length = max(input_ids.shape[-1],negative_ids.shape[-1])
|
168 |
+
|
169 |
+
if input_ids.shape[-1]>negative_ids.shape[-1]:
|
170 |
negative_ids = pipeline.tokenizer(
|
171 |
negative_prompt,
|
172 |
truncation=False,
|
|
|
175 |
return_tensors="pt",
|
176 |
).input_ids.to(device)
|
177 |
else:
|
|
|
|
|
|
|
|
|
178 |
input_ids = pipeline.tokenizer(
|
179 |
+
prompt,
|
180 |
+
return_tensors="pt",
|
181 |
+
truncation=False,
|
182 |
+
padding="max_length",
|
183 |
+
max_length=shape_max_length,
|
184 |
+
).input_ids.to(device)
|
185 |
|
186 |
concat_embeds = []
|
187 |
neg_embeds = []
|
188 |
for i in range(0, shape_max_length, max_length):
|
189 |
+
concat_embeds.append(pipeline.text_encoder(
|
190 |
+
input_ids[:, i: i + max_length])[0])
|
191 |
+
neg_embeds.append(pipeline.text_encoder(
|
192 |
+
negative_ids[:, i: i + max_length])[0])
|
193 |
|
194 |
return torch.cat(concat_embeds, dim=1), torch.cat(neg_embeds, dim=1)
|
195 |
|
|
|
214 |
for layer, elems in updates.items():
|
215 |
|
216 |
if "text" in layer:
|
217 |
+
layer_infos = layer.split(
|
218 |
+
LORA_PREFIX_TEXT_ENCODER + "_")[-1].split("_")
|
219 |
curr_layer = pipeline.text_encoder
|
220 |
else:
|
221 |
+
layer_infos = layer.split(
|
222 |
+
LORA_PREFIX_UNET + "_")[-1].split("_")
|
223 |
curr_layer = pipeline.unet
|
224 |
|
225 |
# find the target layer
|
|
|
282 |
)
|
283 |
curr_layer = pipeline.text_encoder
|
284 |
else:
|
285 |
+
layer_infos = layer.split(
|
286 |
+
LORA_PREFIX_UNET + "_")[-1].split("_")
|
287 |
curr_layer = pipeline.unet
|
288 |
|
289 |
# find the target layer
|
|
|
528 |
self.mask_predictor.set_image(image)
|
529 |
# Separate the points and labels
|
530 |
points, labels = zip(*[(point[:2], point[2])
|
531 |
+
for point in clicked_points])
|
532 |
|
533 |
# Convert the points and labels to numpy arrays
|
534 |
input_point = np.array(points)
|
|
|
573 |
mask_click_np = np.transpose(mask_click_np, (1, 2, 0)) * 255.0
|
574 |
|
575 |
mask_image = HWC3(mask_click_np.astype(np.uint8))
|
576 |
+
mask_image = cv2.resize(
|
577 |
+
mask_image, (W, H), interpolation=cv2.INTER_LINEAR)
|
578 |
# mask_image = Image.fromarray(mask_image_tmp)
|
579 |
|
580 |
# Draw circles for all clicked points
|
|
|
607 |
)
|
608 |
|
609 |
@torch.inference_mode()
|
610 |
+
@save_input_to_file # for debug use
|
611 |
def process(
|
612 |
self,
|
613 |
source_image,
|
|
|
643 |
ref_auto_prompt=False,
|
644 |
ref_textinv=True,
|
645 |
ref_textinv_path=None,
|
646 |
+
ref_scale=None,
|
647 |
):
|
648 |
|
649 |
if condition_model is None or condition_model == "EditAnything":
|
|
|
666 |
)
|
667 |
self.defalut_enable_all_generate = enable_all_generate
|
668 |
if enable_all_generate:
|
|
|
|
|
|
|
|
|
|
|
669 |
mask_image = (
|
670 |
np.ones((input_image.shape[0],
|
671 |
+
input_image.shape[1], 3)) * 255
|
672 |
)
|
673 |
else:
|
674 |
mask_image = source_image["mask"]
|
|
|
736 |
except:
|
737 |
print("No textinvert embeddings found.")
|
738 |
ref_data_path = "./utils/tmp/textinv/img"
|
739 |
+
if not os.path.exists(ref_data_path):
|
740 |
os.makedirs(ref_data_path)
|
741 |
+
cropped_ref_image.save(
|
742 |
+
os.path.join(ref_data_path, 'ref.png'))
|
743 |
print("Ref image region is save to:", ref_data_path)
|
744 |
+
print(
|
745 |
+
"Plese finetune with run_texutal_inversion.sh in utils folder to get the textinvert embeddings.")
|
746 |
|
747 |
else:
|
748 |
ref_mask = None
|
|
|
774 |
)
|
775 |
|
776 |
control = torch.from_numpy(detected_map.copy()).float().cuda()
|
777 |
+
control = control.unsqueeze(dim=0)
|
778 |
control = einops.rearrange(control, "b h w c -> b c h w").clone()
|
779 |
|
780 |
mask_imag_ori = HWC3(mask_image.astype(np.uint8))
|
|
|
792 |
prompt_embeds, negative_prompt_embeds = get_pipeline_embeds(
|
793 |
self.pipe, postive_prompt, negative_prompt, "cuda"
|
794 |
)
|
|
|
|
|
|
|
|
|
795 |
|
796 |
if enable_all_generate and self.extra_inpaint:
|
|
|
|
|
797 |
if ref_image is not None:
|
798 |
print("Not support yet.")
|
799 |
return
|
|
|
878 |
reference_adain=reference_adain,
|
879 |
ref_controlnet_conditioning_scale=ref_multi_condition_scale,
|
880 |
guess_mode=guess_mode,
|
881 |
+
ref_scale=ref_scale,
|
882 |
).images
|
883 |
results = [x_samples[i] for i in range(num_samples)]
|
884 |
|
editany_nogradio.py
ADDED
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pickle
|
2 |
+
from editany_lora import EditAnythingLoraModel
|
3 |
+
model = EditAnythingLoraModel(
|
4 |
+
base_model_path="runwayml/stable-diffusion-v1-5",
|
5 |
+
controlmodel_name='LAION Pretrained(v0-4)-SD15',
|
6 |
+
lora_model_path=None, use_blip=False, extra_inpaint=True,
|
7 |
+
)
|
8 |
+
|
9 |
+
with open('input_data.pkl', 'rb') as f:
|
10 |
+
input_data = pickle.load(f)
|
11 |
+
|
12 |
+
print(input_data)
|
13 |
+
|
14 |
+
refined, output, ref, text = model.process(*input_data['args'], **input_data['kwargs'])
|
15 |
+
|
16 |
+
output
|
17 |
+
|
18 |
+
# a woman in a tan suit and white shirt
|
19 |
+
|
20 |
+
# best quality, extremely detailed,iron man wallpaper
|
editany_test.py
CHANGED
@@ -70,4 +70,4 @@ if __name__ == "__main__":
|
|
70 |
lora_weight=0.5,
|
71 |
)
|
72 |
demo = create_demo(model.process, model.process_image_click)
|
73 |
-
demo.queue().launch(server_name="0.0.0.0")
|
|
|
70 |
lora_weight=0.5,
|
71 |
)
|
72 |
demo = create_demo(model.process, model.process_image_click)
|
73 |
+
demo.queue().launch(server_name="0.0.0.0", share=True)
|
environment.yaml
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
name: control
|
2 |
+
channels:
|
3 |
+
- pytorch
|
4 |
+
- defaults
|
5 |
+
dependencies:
|
6 |
+
- python=3.8.5
|
7 |
+
- pip=20.3
|
8 |
+
- cudatoolkit=11.3
|
9 |
+
- pytorch=1.13.1
|
10 |
+
- torchvision=0.14.1
|
11 |
+
- numpy=1.23.1
|
12 |
+
- pip:
|
13 |
+
- gradio==3.35.2
|
14 |
+
- albumentations==1.3.0
|
15 |
+
- opencv-contrib-python==4.3.0.36
|
16 |
+
- imageio==2.9.0
|
17 |
+
- imageio-ffmpeg==0.4.2
|
18 |
+
- pytorch-lightning==1.5.0
|
19 |
+
- omegaconf==2.1.1
|
20 |
+
- test-tube>=0.7.5
|
21 |
+
- streamlit==1.12.1
|
22 |
+
- einops==0.3.0
|
23 |
+
- webdataset==0.2.5
|
24 |
+
- kornia==0.6
|
25 |
+
- open_clip_torch==2.0.2
|
26 |
+
- invisible-watermark>=0.1.5
|
27 |
+
- streamlit-drawable-canvas==0.8.0
|
28 |
+
- torchmetrics==0.6.0
|
29 |
+
- timm==0.6.12
|
30 |
+
- addict==2.4.0
|
31 |
+
- yapf==0.32.0
|
32 |
+
- prettytable==3.6.0
|
33 |
+
- safetensors==0.2.7
|
34 |
+
- basicsr==1.4.2
|
35 |
+
- diffusers==0.17.1
|
36 |
+
- accelerate==0.17.0
|
37 |
+
- transformers==4.30.2
|
38 |
+
- xformers
|
requirements.txt
CHANGED
@@ -30,4 +30,4 @@ transformers==4.30.2
|
|
30 |
xformers==0.0.16
|
31 |
triton
|
32 |
gradio==3.35.2
|
33 |
-
gradio-client==0.2.7
|
|
|
30 |
xformers==0.0.16
|
31 |
triton
|
32 |
gradio==3.35.2
|
33 |
+
gradio-client==0.2.7
|
utils/stable_diffusion_controlnet_inpaint.py
CHANGED
@@ -1179,6 +1179,7 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1179 |
style_fidelity: float = 0.5,
|
1180 |
reference_attn: bool = True,
|
1181 |
reference_adain: bool = True,
|
|
|
1182 |
):
|
1183 |
r"""
|
1184 |
Function invoked when calling the pipeline for generation.
|
@@ -1272,6 +1273,8 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1272 |
Whether to use reference query for self attention's context.
|
1273 |
reference_adain (`bool`):
|
1274 |
Whether to use reference adain.
|
|
|
|
|
1275 |
|
1276 |
Examples:
|
1277 |
|
@@ -1346,8 +1349,9 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1346 |
ref_prompt_embeds = self._encode_prompt(
|
1347 |
ref_prompt,
|
1348 |
device,
|
1349 |
-
num_images_per_prompt * 2,
|
1350 |
-
|
|
|
1351 |
negative_prompt="longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality",
|
1352 |
prompt_embeds=None,
|
1353 |
)
|
@@ -1414,13 +1418,13 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1414 |
num_images_per_prompt=num_images_per_prompt,
|
1415 |
device=device,
|
1416 |
dtype=self.controlnet.dtype,
|
1417 |
-
do_classifier_free_guidance=
|
1418 |
)
|
1419 |
ref_controlnet_conditioning_image = controlnet_conditioning_image.copy()
|
|
|
|
|
|
|
1420 |
ref_controlnet_conditioning_image[-1] = ref_control_image
|
1421 |
-
# ref_controlnet_conditioning_scale = controlnet_conditioning_scale.copy()
|
1422 |
-
# ref_controlnet_conditioning_scale[0] = 1.0 # disable the first sam controlnet
|
1423 |
-
# ref_controlnet_conditioning_scale[-1] = 0.2
|
1424 |
|
1425 |
# 5. Prepare timesteps
|
1426 |
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
@@ -1491,7 +1495,7 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1491 |
prompt_embeds.dtype,
|
1492 |
device,
|
1493 |
generator,
|
1494 |
-
|
1495 |
)
|
1496 |
|
1497 |
# 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
@@ -1511,6 +1515,7 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1511 |
self.gn_auto_machine_weight = gn_auto_machine_weight
|
1512 |
self.do_classifier_free_guidance = do_classifier_free_guidance
|
1513 |
self.style_fidelity = style_fidelity
|
|
|
1514 |
self.ref_mask = ref_mask
|
1515 |
self.inpaint_mask = mask_image
|
1516 |
attn_modules, gn_modules = self.redefine_ref_model(
|
@@ -1518,9 +1523,16 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1518 |
)
|
1519 |
|
1520 |
control_attn_modules, control_gn_modules = self.redefine_ref_model(
|
1521 |
-
self.controlnet, reference_attn,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1522 |
)
|
1523 |
-
|
1524 |
# 8. Denoising loop
|
1525 |
num_warmup_steps = len(timesteps) - \
|
1526 |
num_inference_steps * self.scheduler.order
|
@@ -1549,12 +1561,6 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1549 |
|
1550 |
if ref_image is not None: # for ref_only mode
|
1551 |
# ref only part
|
1552 |
-
noise = randn_tensor(
|
1553 |
-
ref_image_latents.shape,
|
1554 |
-
generator=generator,
|
1555 |
-
device=ref_image_latents.device,
|
1556 |
-
dtype=ref_image_latents.dtype,
|
1557 |
-
)
|
1558 |
ref_xt = self.scheduler.add_noise(
|
1559 |
ref_image_latents,
|
1560 |
noise,
|
@@ -1566,8 +1572,8 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1566 |
|
1567 |
MODE = "write"
|
1568 |
self.change_module_mode(
|
1569 |
-
MODE, control_attn_modules, control_gn_modules
|
1570 |
-
)
|
1571 |
|
1572 |
(
|
1573 |
ref_down_block_res_samples,
|
@@ -1582,7 +1588,6 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1582 |
return_dict=False,
|
1583 |
)
|
1584 |
|
1585 |
-
self.change_module_mode(MODE, attn_modules, gn_modules)
|
1586 |
self.unet(
|
1587 |
ref_xt,
|
1588 |
t,
|
@@ -1595,7 +1600,10 @@ class StableDiffusionControlNetInpaintPipeline(
|
|
1595 |
|
1596 |
# predict the noise residual
|
1597 |
MODE = "read" # change to read mode for following noise_pred
|
|
|
|
|
1598 |
self.change_module_mode(MODE, attn_modules, gn_modules)
|
|
|
1599 |
down_block_res_samples, mid_block_res_sample = self.controlnet(
|
1600 |
non_inpainting_latent_model_input,
|
1601 |
t,
|
|
|
1179 |
style_fidelity: float = 0.5,
|
1180 |
reference_attn: bool = True,
|
1181 |
reference_adain: bool = True,
|
1182 |
+
ref_scale: float = 1.0,
|
1183 |
):
|
1184 |
r"""
|
1185 |
Function invoked when calling the pipeline for generation.
|
|
|
1273 |
Whether to use reference query for self attention's context.
|
1274 |
reference_adain (`bool`):
|
1275 |
Whether to use reference adain.
|
1276 |
+
ref_scale (`float`):
|
1277 |
+
reference guidance scale.
|
1278 |
|
1279 |
Examples:
|
1280 |
|
|
|
1349 |
ref_prompt_embeds = self._encode_prompt(
|
1350 |
ref_prompt,
|
1351 |
device,
|
1352 |
+
# num_images_per_prompt * 2,
|
1353 |
+
num_images_per_prompt * 1,
|
1354 |
+
False,
|
1355 |
negative_prompt="longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality",
|
1356 |
prompt_embeds=None,
|
1357 |
)
|
|
|
1418 |
num_images_per_prompt=num_images_per_prompt,
|
1419 |
device=device,
|
1420 |
dtype=self.controlnet.dtype,
|
1421 |
+
do_classifier_free_guidance=False,
|
1422 |
)
|
1423 |
ref_controlnet_conditioning_image = controlnet_conditioning_image.copy()
|
1424 |
+
for i in range(len(ref_controlnet_conditioning_image)):
|
1425 |
+
ref_controlnet_conditioning_image[i] = ref_controlnet_conditioning_image[i].chunk(
|
1426 |
+
2)[0] # remove the extra guidance for cfg
|
1427 |
ref_controlnet_conditioning_image[-1] = ref_control_image
|
|
|
|
|
|
|
1428 |
|
1429 |
# 5. Prepare timesteps
|
1430 |
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
|
|
1495 |
prompt_embeds.dtype,
|
1496 |
device,
|
1497 |
generator,
|
1498 |
+
False,
|
1499 |
)
|
1500 |
|
1501 |
# 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
|
|
1515 |
self.gn_auto_machine_weight = gn_auto_machine_weight
|
1516 |
self.do_classifier_free_guidance = do_classifier_free_guidance
|
1517 |
self.style_fidelity = style_fidelity
|
1518 |
+
self.ref_scale = ref_scale
|
1519 |
self.ref_mask = ref_mask
|
1520 |
self.inpaint_mask = mask_image
|
1521 |
attn_modules, gn_modules = self.redefine_ref_model(
|
|
|
1523 |
)
|
1524 |
|
1525 |
control_attn_modules, control_gn_modules = self.redefine_ref_model(
|
1526 |
+
self.controlnet, reference_attn, reference_adain, model_type="controlnet"
|
1527 |
+
)
|
1528 |
+
if ref_image is not None:
|
1529 |
+
noise = randn_tensor(
|
1530 |
+
# ref_image_latents.shape,
|
1531 |
+
latents.shape,
|
1532 |
+
generator=generator,
|
1533 |
+
device=ref_image_latents.device,
|
1534 |
+
dtype=ref_image_latents.dtype,
|
1535 |
)
|
|
|
1536 |
# 8. Denoising loop
|
1537 |
num_warmup_steps = len(timesteps) - \
|
1538 |
num_inference_steps * self.scheduler.order
|
|
|
1561 |
|
1562 |
if ref_image is not None: # for ref_only mode
|
1563 |
# ref only part
|
|
|
|
|
|
|
|
|
|
|
|
|
1564 |
ref_xt = self.scheduler.add_noise(
|
1565 |
ref_image_latents,
|
1566 |
noise,
|
|
|
1572 |
|
1573 |
MODE = "write"
|
1574 |
self.change_module_mode(
|
1575 |
+
MODE, control_attn_modules, control_gn_modules)
|
1576 |
+
self.change_module_mode(MODE, attn_modules, gn_modules)
|
1577 |
|
1578 |
(
|
1579 |
ref_down_block_res_samples,
|
|
|
1588 |
return_dict=False,
|
1589 |
)
|
1590 |
|
|
|
1591 |
self.unet(
|
1592 |
ref_xt,
|
1593 |
t,
|
|
|
1600 |
|
1601 |
# predict the noise residual
|
1602 |
MODE = "read" # change to read mode for following noise_pred
|
1603 |
+
self.change_module_mode(
|
1604 |
+
MODE, control_attn_modules, control_gn_modules)
|
1605 |
self.change_module_mode(MODE, attn_modules, gn_modules)
|
1606 |
+
|
1607 |
down_block_res_samples, mid_block_res_sample = self.controlnet(
|
1608 |
non_inpainting_latent_model_input,
|
1609 |
t,
|
utils/stable_diffusion_reference.py
CHANGED
@@ -1,12 +1,12 @@
|
|
1 |
# Based on https://raw.githubusercontent.com/okotaku/diffusers/feature/reference_only_control/examples/community/stable_diffusion_reference.py
|
2 |
# Inspired by: https://github.com/Mikubill/sd-webui-controlnet/discussions/1236
|
|
|
3 |
from typing import Any, Callable, Dict, List, Optional, Union, Tuple
|
4 |
|
5 |
import numpy as np
|
6 |
import PIL.Image
|
7 |
import torch
|
8 |
|
9 |
-
from diffusers import StableDiffusionPipeline
|
10 |
from diffusers.models.attention import BasicTransformerBlock
|
11 |
from diffusers.models.unet_2d_blocks import (
|
12 |
CrossAttnDownBlock2D,
|
@@ -14,11 +14,9 @@ from diffusers.models.unet_2d_blocks import (
|
|
14 |
DownBlock2D,
|
15 |
UpBlock2D,
|
16 |
)
|
17 |
-
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
|
18 |
from diffusers.utils import PIL_INTERPOLATION, logging
|
19 |
import torch.nn.functional as F
|
20 |
|
21 |
-
|
22 |
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
23 |
|
24 |
EXAMPLE_DOC_STRING = """
|
@@ -56,6 +54,127 @@ def torch_dfs(model: torch.nn.Module):
|
|
56 |
return result
|
57 |
|
58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
class StableDiffusionReferencePipeline:
|
60 |
def prepare_ref_image(
|
61 |
self,
|
@@ -237,9 +356,8 @@ class StableDiffusionReferencePipeline:
|
|
237 |
this_ref_mask = F.interpolate(
|
238 |
this_ref_mask, scale_factor=ref_scale
|
239 |
)
|
240 |
-
|
241 |
-
|
242 |
-
# this_ref_mask = this_ref_mask.view(1,-1,1)
|
243 |
this_ref_mask = this_ref_mask.repeat(
|
244 |
resize_norm_hidden_states.shape[0],
|
245 |
resize_norm_hidden_states.shape[1],
|
@@ -256,11 +374,14 @@ class StableDiffusionReferencePipeline:
|
|
256 |
-1,
|
257 |
)
|
258 |
)
|
|
|
259 |
masked_norm_hidden_states = masked_norm_hidden_states.permute(
|
260 |
0, 2, 1
|
261 |
)
|
262 |
self.bank.append(masked_norm_hidden_states)
|
263 |
-
|
|
|
|
|
264 |
attn_output = self.attn1(
|
265 |
norm_hidden_states,
|
266 |
encoder_hidden_states=encoder_hidden_states
|
@@ -271,31 +392,27 @@ class StableDiffusionReferencePipeline:
|
|
271 |
)
|
272 |
if self.MODE == "read":
|
273 |
if self.attention_auto_machine_weight > self.attn_weight:
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
# this_ref_mask = this_ref_mask.repeat(norm_hidden_states.shape[0], norm_hidden_states.shape[1], 1)
|
284 |
-
# this_ref_mask = torch.zeros(
|
285 |
-
# norm_hidden_states.shape[0], norm_hidden_states.shape[1], this_ref_mask.shape[1], dtype=norm_hidden_states.dtype, device=norm_hidden_states.device
|
286 |
-
# )
|
287 |
-
# print(attention_mask.shape, this_ref_mask.shape)
|
288 |
-
# attention_mask = torch.cat((attention_mask, this_ref_mask), dim=-1)
|
289 |
-
# print("merge", attention_mask.shape)
|
290 |
ref_hidden_states = torch.cat(
|
291 |
-
|
292 |
)
|
|
|
|
|
|
|
293 |
attn_output_uc = self.attn1(
|
294 |
-
|
295 |
encoder_hidden_states=ref_hidden_states,
|
296 |
-
# attention_mask=attention_mask,
|
297 |
**cross_attention_kwargs,
|
298 |
)
|
|
|
299 |
attn_output_c = attn_output_uc.clone()
|
300 |
if self.do_classifier_free_guidance and self.style_fidelity > 0:
|
301 |
attn_output_c[self.uc_mask] = self.attn1(
|
@@ -308,6 +425,9 @@ class StableDiffusionReferencePipeline:
|
|
308 |
+ (1.0 - self.style_fidelity) * attn_output_uc
|
309 |
)
|
310 |
self.bank.clear()
|
|
|
|
|
|
|
311 |
else:
|
312 |
attn_output = self.attn1(
|
313 |
norm_hidden_states,
|
@@ -317,6 +437,9 @@ class StableDiffusionReferencePipeline:
|
|
317 |
attention_mask=attention_mask,
|
318 |
**cross_attention_kwargs,
|
319 |
)
|
|
|
|
|
|
|
320 |
if self.use_ada_layer_norm_zero:
|
321 |
attn_output = gate_msa.unsqueeze(1) * attn_output
|
322 |
hidden_states = attn_output + hidden_states
|
@@ -365,6 +488,10 @@ class StableDiffusionReferencePipeline:
|
|
365 |
this_ref_mask = F.interpolate(
|
366 |
self.ref_mask.to(x.device), scale_factor=1 / scale_ratio
|
367 |
)
|
|
|
|
|
|
|
|
|
368 |
this_ref_mask = this_ref_mask.repeat(
|
369 |
x.shape[0], x.shape[1], 1, 1
|
370 |
).bool()
|
@@ -378,8 +505,8 @@ class StableDiffusionReferencePipeline:
|
|
378 |
masked_x, dim=(2, 3), keepdim=True, correction=0
|
379 |
)
|
380 |
|
381 |
-
self.mean_bank.append(mean)
|
382 |
-
self.var_bank.append(var)
|
383 |
if self.MODE == "read":
|
384 |
if (
|
385 |
self.gn_auto_machine_weight >= self.gn_weight
|
@@ -387,37 +514,12 @@ class StableDiffusionReferencePipeline:
|
|
387 |
and len(self.var_bank) > 0
|
388 |
):
|
389 |
# print("hacked_mid_forward")
|
390 |
-
|
391 |
-
|
392 |
-
|
393 |
-
|
394 |
-
|
395 |
-
|
396 |
-
).bool()
|
397 |
-
masked_x = (
|
398 |
-
x[this_inpaint_mask]
|
399 |
-
.detach()
|
400 |
-
.clone()
|
401 |
-
.view(x.shape[0], x.shape[1], -1, 1)
|
402 |
-
)
|
403 |
-
var, mean = torch.var_mean(
|
404 |
-
masked_x, dim=(2, 3), keepdim=True, correction=0
|
405 |
-
)
|
406 |
-
std = torch.maximum(
|
407 |
-
var, torch.zeros_like(var) + eps) ** 0.5
|
408 |
-
mean_acc = sum(self.mean_bank) / float(len(self.mean_bank))
|
409 |
-
var_acc = sum(self.var_bank) / float(len(self.var_bank))
|
410 |
-
std_acc = (
|
411 |
-
torch.maximum(var_acc, torch.zeros_like(
|
412 |
-
var_acc) + eps) ** 0.5
|
413 |
-
)
|
414 |
-
x_uc = (((masked_x - mean) / std) * std_acc) + mean_acc
|
415 |
-
x_c = x_uc.clone()
|
416 |
-
if self.do_classifier_free_guidance and self.style_fidelity > 0:
|
417 |
-
x_c[self.uc_mask] = masked_x[self.uc_mask]
|
418 |
-
masked_x = self.style_fidelity * x_c + \
|
419 |
-
(1.0 - self.style_fidelity) * x_uc
|
420 |
-
x[this_inpaint_mask] = masked_x.view(-1)
|
421 |
self.mean_bank = []
|
422 |
self.var_bank = []
|
423 |
return x
|
@@ -448,6 +550,8 @@ class StableDiffusionReferencePipeline:
|
|
448 |
self.ref_mask.to(hidden_states.device),
|
449 |
scale_factor=1 / scale_ratio,
|
450 |
)
|
|
|
|
|
451 |
this_ref_mask = this_ref_mask.repeat(
|
452 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
453 |
).bool()
|
@@ -460,8 +564,8 @@ class StableDiffusionReferencePipeline:
|
|
460 |
var, mean = torch.var_mean(
|
461 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
462 |
)
|
463 |
-
self.mean_bank0.append(mean)
|
464 |
-
self.var_bank0.append(var)
|
465 |
if self.MODE == "read":
|
466 |
if (
|
467 |
self.gn_auto_machine_weight >= self.gn_weight
|
@@ -469,54 +573,17 @@ class StableDiffusionReferencePipeline:
|
|
469 |
and len(self.var_bank0) > 0
|
470 |
):
|
471 |
# print("hacked_CrossAttnDownBlock2D_forward0")
|
472 |
-
|
473 |
-
hidden_states.
|
474 |
-
|
475 |
-
|
476 |
-
|
477 |
-
|
478 |
-
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
479 |
-
).bool()
|
480 |
-
masked_hidden_states = (
|
481 |
-
hidden_states[this_inpaint_mask]
|
482 |
-
.detach()
|
483 |
-
.clone()
|
484 |
-
.view(hidden_states.shape[0], hidden_states.shape[1], -1, 1)
|
485 |
-
)
|
486 |
-
var, mean = torch.var_mean(
|
487 |
-
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
488 |
-
)
|
489 |
-
std = torch.maximum(
|
490 |
-
var, torch.zeros_like(var) + eps) ** 0.5
|
491 |
-
mean_acc = sum(self.mean_bank0[i]) / float(
|
492 |
-
len(self.mean_bank0[i])
|
493 |
-
)
|
494 |
-
var_acc = sum(
|
495 |
-
self.var_bank0[i]) / float(len(self.var_bank0[i]))
|
496 |
-
std_acc = (
|
497 |
-
torch.maximum(
|
498 |
-
var_acc, torch.zeros_like(var_acc) + eps)
|
499 |
-
** 0.5
|
500 |
-
)
|
501 |
-
hidden_states_uc = (
|
502 |
-
((masked_hidden_states - mean) / std) * std_acc
|
503 |
-
) + mean_acc
|
504 |
-
hidden_states_c = hidden_states_uc.clone()
|
505 |
-
if self.do_classifier_free_guidance and self.style_fidelity > 0:
|
506 |
-
hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask]
|
507 |
-
masked_hidden_states = (
|
508 |
-
self.style_fidelity * hidden_states_c
|
509 |
-
+ (1.0 - self.style_fidelity) * hidden_states_uc
|
510 |
-
)
|
511 |
-
hidden_states[this_inpaint_mask] = masked_hidden_states.view(
|
512 |
-
-1)
|
513 |
|
514 |
hidden_states = attn(
|
515 |
hidden_states,
|
516 |
encoder_hidden_states=encoder_hidden_states,
|
517 |
cross_attention_kwargs=cross_attention_kwargs,
|
518 |
-
# attention_mask=attention_mask,
|
519 |
-
# encoder_attention_mask=encoder_attention_mask,
|
520 |
return_dict=False,
|
521 |
)[0]
|
522 |
if self.MODE == "write":
|
@@ -528,6 +595,8 @@ class StableDiffusionReferencePipeline:
|
|
528 |
self.ref_mask.to(hidden_states.device),
|
529 |
scale_factor=1 / scale_ratio,
|
530 |
)
|
|
|
|
|
531 |
this_ref_mask = this_ref_mask.repeat(
|
532 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
533 |
).bool()
|
@@ -540,8 +609,8 @@ class StableDiffusionReferencePipeline:
|
|
540 |
var, mean = torch.var_mean(
|
541 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
542 |
)
|
543 |
-
self.mean_bank.append(mean)
|
544 |
-
self.var_bank.append(var)
|
545 |
if self.MODE == "read":
|
546 |
if (
|
547 |
self.gn_auto_machine_weight >= self.gn_weight
|
@@ -549,48 +618,12 @@ class StableDiffusionReferencePipeline:
|
|
549 |
and len(self.var_bank) > 0
|
550 |
):
|
551 |
# print("hack_CrossAttnDownBlock2D_forward")
|
552 |
-
|
553 |
-
hidden_states.
|
554 |
-
this_inpaint_mask = F.interpolate(
|
555 |
-
self.inpaint_mask.to(hidden_states.device), scale_factor=1 / scale_ratio
|
556 |
-
)
|
557 |
-
this_inpaint_mask = this_inpaint_mask.repeat(
|
558 |
-
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
559 |
-
).bool()
|
560 |
-
masked_hidden_states = (
|
561 |
-
hidden_states[this_inpaint_mask]
|
562 |
-
.detach()
|
563 |
-
.clone()
|
564 |
-
.view(hidden_states.shape[0], hidden_states.shape[1], -1, 1)
|
565 |
-
)
|
566 |
-
var, mean = torch.var_mean(
|
567 |
-
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
568 |
-
)
|
569 |
-
std = torch.maximum(
|
570 |
-
var, torch.zeros_like(var) + eps) ** 0.5
|
571 |
-
mean_acc = sum(self.mean_bank[i]) / float(
|
572 |
-
len(self.mean_bank[i])
|
573 |
-
)
|
574 |
-
var_acc = sum(
|
575 |
-
self.var_bank[i]) / float(len(self.var_bank[i]))
|
576 |
-
std_acc = (
|
577 |
-
torch.maximum(
|
578 |
-
var_acc, torch.zeros_like(var_acc) + eps)
|
579 |
-
** 0.5
|
580 |
-
)
|
581 |
-
hidden_states_uc = (
|
582 |
-
((masked_hidden_states - mean) / std) * std_acc
|
583 |
-
) + mean_acc
|
584 |
-
hidden_states_c = hidden_states_uc.clone()
|
585 |
-
if self.do_classifier_free_guidance and self.style_fidelity > 0:
|
586 |
-
hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask]
|
587 |
-
masked_hidden_states = (
|
588 |
-
self.style_fidelity * hidden_states_c
|
589 |
-
+ (1.0 - self.style_fidelity) * hidden_states_uc
|
590 |
-
)
|
591 |
-
hidden_states[this_inpaint_mask] = masked_hidden_states.view(
|
592 |
-
-1)
|
593 |
|
|
|
|
|
|
|
594 |
output_states = output_states + (hidden_states,)
|
595 |
|
596 |
if self.MODE == "read":
|
@@ -598,6 +631,8 @@ class StableDiffusionReferencePipeline:
|
|
598 |
self.var_bank0 = []
|
599 |
self.mean_bank = []
|
600 |
self.var_bank = []
|
|
|
|
|
601 |
|
602 |
if self.downsamplers is not None:
|
603 |
for downsampler in self.downsamplers:
|
@@ -625,6 +660,8 @@ class StableDiffusionReferencePipeline:
|
|
625 |
self.ref_mask.to(hidden_states.device),
|
626 |
scale_factor=1 / scale_ratio,
|
627 |
)
|
|
|
|
|
628 |
this_ref_mask = this_ref_mask.repeat(
|
629 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
630 |
).bool()
|
@@ -637,8 +674,8 @@ class StableDiffusionReferencePipeline:
|
|
637 |
var, mean = torch.var_mean(
|
638 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
639 |
)
|
640 |
-
self.mean_bank.append(mean)
|
641 |
-
self.var_bank.append(var)
|
642 |
if self.MODE == "read":
|
643 |
if (
|
644 |
self.gn_auto_machine_weight >= self.gn_weight
|
@@ -646,53 +683,19 @@ class StableDiffusionReferencePipeline:
|
|
646 |
and len(self.var_bank) > 0
|
647 |
):
|
648 |
# print("hacked_DownBlock2D_forward")
|
649 |
-
|
650 |
-
hidden_states.
|
651 |
-
|
652 |
-
|
653 |
-
|
654 |
-
|
655 |
-
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
656 |
-
).bool()
|
657 |
-
masked_hidden_states = (
|
658 |
-
hidden_states[this_inpaint_mask]
|
659 |
-
.detach()
|
660 |
-
.clone()
|
661 |
-
.view(hidden_states.shape[0], hidden_states.shape[1], -1, 1)
|
662 |
-
)
|
663 |
-
var, mean = torch.var_mean(
|
664 |
-
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
665 |
-
)
|
666 |
-
std = torch.maximum(
|
667 |
-
var, torch.zeros_like(var) + eps) ** 0.5
|
668 |
-
mean_acc = sum(self.mean_bank[i]) / float(
|
669 |
-
len(self.mean_bank[i])
|
670 |
-
)
|
671 |
-
var_acc = sum(
|
672 |
-
self.var_bank[i]) / float(len(self.var_bank[i]))
|
673 |
-
std_acc = (
|
674 |
-
torch.maximum(
|
675 |
-
var_acc, torch.zeros_like(var_acc) + eps)
|
676 |
-
** 0.5
|
677 |
-
)
|
678 |
-
hidden_states_uc = (
|
679 |
-
((masked_hidden_states - mean) / std) * std_acc
|
680 |
-
) + mean_acc
|
681 |
-
hidden_states_c = hidden_states_uc.clone()
|
682 |
-
if self.do_classifier_free_guidance and self.style_fidelity > 0:
|
683 |
-
hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask]
|
684 |
-
masked_hidden_states = (
|
685 |
-
self.style_fidelity * hidden_states_c
|
686 |
-
+ (1.0 - self.style_fidelity) * hidden_states_uc
|
687 |
-
)
|
688 |
-
hidden_states[this_inpaint_mask] = masked_hidden_states.view(
|
689 |
-
-1)
|
690 |
|
691 |
output_states = output_states + (hidden_states,)
|
692 |
|
693 |
if self.MODE == "read":
|
694 |
self.mean_bank = []
|
695 |
self.var_bank = []
|
|
|
696 |
|
697 |
if self.downsamplers is not None:
|
698 |
for downsampler in self.downsamplers:
|
@@ -733,6 +736,8 @@ class StableDiffusionReferencePipeline:
|
|
733 |
self.ref_mask.to(hidden_states.device),
|
734 |
scale_factor=1 / scale_ratio,
|
735 |
)
|
|
|
|
|
736 |
this_ref_mask = this_ref_mask.repeat(
|
737 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
738 |
).bool()
|
@@ -745,8 +750,8 @@ class StableDiffusionReferencePipeline:
|
|
745 |
var, mean = torch.var_mean(
|
746 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
747 |
)
|
748 |
-
self.mean_bank0.append(mean)
|
749 |
-
self.var_bank0.append(var)
|
750 |
if self.MODE == "read":
|
751 |
if (
|
752 |
self.gn_auto_machine_weight >= self.gn_weight
|
@@ -754,47 +759,12 @@ class StableDiffusionReferencePipeline:
|
|
754 |
and len(self.var_bank0) > 0
|
755 |
):
|
756 |
# print("hacked_CrossAttnUpBlock2D_forward1")
|
757 |
-
|
758 |
-
hidden_states.
|
759 |
-
|
760 |
-
|
761 |
-
|
762 |
-
|
763 |
-
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
764 |
-
).bool()
|
765 |
-
masked_hidden_states = (
|
766 |
-
hidden_states[this_inpaint_mask]
|
767 |
-
.detach()
|
768 |
-
.clone()
|
769 |
-
.view(hidden_states.shape[0], hidden_states.shape[1], -1, 1)
|
770 |
-
)
|
771 |
-
var, mean = torch.var_mean(
|
772 |
-
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
773 |
-
)
|
774 |
-
std = torch.maximum(
|
775 |
-
var, torch.zeros_like(var) + eps) ** 0.5
|
776 |
-
mean_acc = sum(self.mean_bank0[i]) / float(
|
777 |
-
len(self.mean_bank0[i])
|
778 |
-
)
|
779 |
-
var_acc = sum(
|
780 |
-
self.var_bank0[i]) / float(len(self.var_bank0[i]))
|
781 |
-
std_acc = (
|
782 |
-
torch.maximum(
|
783 |
-
var_acc, torch.zeros_like(var_acc) + eps)
|
784 |
-
** 0.5
|
785 |
-
)
|
786 |
-
hidden_states_uc = (
|
787 |
-
((masked_hidden_states - mean) / std) * std_acc
|
788 |
-
) + mean_acc
|
789 |
-
hidden_states_c = hidden_states_uc.clone()
|
790 |
-
if self.do_classifier_free_guidance and self.style_fidelity > 0:
|
791 |
-
hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask]
|
792 |
-
masked_hidden_states = (
|
793 |
-
self.style_fidelity * hidden_states_c
|
794 |
-
+ (1.0 - self.style_fidelity) * hidden_states_uc
|
795 |
-
)
|
796 |
-
hidden_states[this_inpaint_mask] = masked_hidden_states.view(
|
797 |
-
-1)
|
798 |
|
799 |
hidden_states = attn(
|
800 |
hidden_states,
|
@@ -815,6 +785,8 @@ class StableDiffusionReferencePipeline:
|
|
815 |
self.ref_mask.to(hidden_states.device),
|
816 |
scale_factor=1 / scale_ratio,
|
817 |
)
|
|
|
|
|
818 |
this_ref_mask = this_ref_mask.repeat(
|
819 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
820 |
).bool()
|
@@ -827,8 +799,8 @@ class StableDiffusionReferencePipeline:
|
|
827 |
var, mean = torch.var_mean(
|
828 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
829 |
)
|
830 |
-
self.mean_bank.append(mean)
|
831 |
-
self.var_bank.append(var)
|
832 |
if self.MODE == "read":
|
833 |
if (
|
834 |
self.gn_auto_machine_weight >= self.gn_weight
|
@@ -836,53 +808,20 @@ class StableDiffusionReferencePipeline:
|
|
836 |
and len(self.var_bank) > 0
|
837 |
):
|
838 |
# print("hacked_CrossAttnUpBlock2D_forward")
|
839 |
-
|
840 |
-
hidden_states.
|
841 |
-
|
842 |
-
|
843 |
-
|
844 |
-
|
845 |
-
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
846 |
-
).bool()
|
847 |
-
masked_hidden_states = (
|
848 |
-
hidden_states[this_inpaint_mask]
|
849 |
-
.detach()
|
850 |
-
.clone()
|
851 |
-
.view(hidden_states.shape[0], hidden_states.shape[1], -1, 1)
|
852 |
-
)
|
853 |
-
var, mean = torch.var_mean(
|
854 |
-
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
855 |
-
)
|
856 |
-
std = torch.maximum(
|
857 |
-
var, torch.zeros_like(var) + eps) ** 0.5
|
858 |
-
mean_acc = sum(self.mean_bank[i]) / float(
|
859 |
-
len(self.mean_bank[i])
|
860 |
-
)
|
861 |
-
var_acc = sum(
|
862 |
-
self.var_bank[i]) / float(len(self.var_bank[i]))
|
863 |
-
std_acc = (
|
864 |
-
torch.maximum(
|
865 |
-
var_acc, torch.zeros_like(var_acc) + eps)
|
866 |
-
** 0.5
|
867 |
-
)
|
868 |
-
hidden_states_uc = (
|
869 |
-
((masked_hidden_states - mean) / std) * std_acc
|
870 |
-
) + mean_acc
|
871 |
-
hidden_states_c = hidden_states_uc.clone()
|
872 |
-
if self.do_classifier_free_guidance and self.style_fidelity > 0:
|
873 |
-
hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask]
|
874 |
-
masked_hidden_states = (
|
875 |
-
self.style_fidelity * hidden_states_c
|
876 |
-
+ (1.0 - self.style_fidelity) * hidden_states_uc
|
877 |
-
)
|
878 |
-
hidden_states[this_inpaint_mask] = masked_hidden_states.view(
|
879 |
-
-1)
|
880 |
|
881 |
if self.MODE == "read":
|
882 |
self.mean_bank0 = []
|
883 |
self.var_bank0 = []
|
884 |
self.mean_bank = []
|
885 |
self.var_bank = []
|
|
|
|
|
886 |
|
887 |
if self.upsamplers is not None:
|
888 |
for upsampler in self.upsamplers:
|
@@ -912,6 +851,8 @@ class StableDiffusionReferencePipeline:
|
|
912 |
self.ref_mask.to(hidden_states.device),
|
913 |
scale_factor=1 / scale_ratio,
|
914 |
)
|
|
|
|
|
915 |
this_ref_mask = this_ref_mask.repeat(
|
916 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
917 |
).bool()
|
@@ -924,8 +865,8 @@ class StableDiffusionReferencePipeline:
|
|
924 |
var, mean = torch.var_mean(
|
925 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
926 |
)
|
927 |
-
self.mean_bank.append(mean)
|
928 |
-
self.var_bank.append(var)
|
929 |
if self.MODE == "read":
|
930 |
if (
|
931 |
self.gn_auto_machine_weight >= self.gn_weight
|
@@ -933,51 +874,17 @@ class StableDiffusionReferencePipeline:
|
|
933 |
and len(self.var_bank) > 0
|
934 |
):
|
935 |
# print("hacked_UpBlock2D_forward")
|
936 |
-
|
937 |
-
hidden_states.
|
938 |
-
|
939 |
-
|
940 |
-
|
941 |
-
|
942 |
-
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
943 |
-
).bool()
|
944 |
-
masked_hidden_states = (
|
945 |
-
hidden_states[this_inpaint_mask]
|
946 |
-
.detach()
|
947 |
-
.clone()
|
948 |
-
.view(hidden_states.shape[0], hidden_states.shape[1], -1, 1)
|
949 |
-
)
|
950 |
-
var, mean = torch.var_mean(
|
951 |
-
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
952 |
-
)
|
953 |
-
std = torch.maximum(
|
954 |
-
var, torch.zeros_like(var) + eps) ** 0.5
|
955 |
-
mean_acc = sum(self.mean_bank[i]) / float(
|
956 |
-
len(self.mean_bank[i])
|
957 |
-
)
|
958 |
-
var_acc = sum(
|
959 |
-
self.var_bank[i]) / float(len(self.var_bank[i]))
|
960 |
-
std_acc = (
|
961 |
-
torch.maximum(
|
962 |
-
var_acc, torch.zeros_like(var_acc) + eps)
|
963 |
-
** 0.5
|
964 |
-
)
|
965 |
-
hidden_states_uc = (
|
966 |
-
((masked_hidden_states - mean) / std) * std_acc
|
967 |
-
) + mean_acc
|
968 |
-
hidden_states_c = hidden_states_uc.clone()
|
969 |
-
if self.do_classifier_free_guidance and self.style_fidelity > 0:
|
970 |
-
hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask]
|
971 |
-
masked_hidden_states = (
|
972 |
-
self.style_fidelity * hidden_states_c
|
973 |
-
+ (1.0 - self.style_fidelity) * hidden_states_uc
|
974 |
-
)
|
975 |
-
hidden_states[this_inpaint_mask] = masked_hidden_states.view(
|
976 |
-
-1)
|
977 |
|
978 |
if self.MODE == "read":
|
979 |
self.mean_bank = []
|
980 |
self.var_bank = []
|
|
|
981 |
|
982 |
if self.upsamplers is not None:
|
983 |
for upsampler in self.upsamplers:
|
@@ -1003,6 +910,7 @@ class StableDiffusionReferencePipeline:
|
|
1003 |
module, BasicTransformerBlock
|
1004 |
)
|
1005 |
module.bank = []
|
|
|
1006 |
module.attn_weight = float(i) / float(len(attn_modules))
|
1007 |
module.attention_auto_machine_weight = (
|
1008 |
self.attention_auto_machine_weight
|
@@ -1017,6 +925,7 @@ class StableDiffusionReferencePipeline:
|
|
1017 |
module.uc_mask = self.uc_mask
|
1018 |
module.style_fidelity = self.style_fidelity
|
1019 |
module.ref_mask = self.ref_mask
|
|
|
1020 |
else:
|
1021 |
attn_modules = None
|
1022 |
if reference_adain:
|
@@ -1043,12 +952,14 @@ class StableDiffusionReferencePipeline:
|
|
1043 |
module.forward = hacked_mid_forward.__get__(
|
1044 |
module, torch.nn.Module
|
1045 |
)
|
1046 |
-
elif isinstance(module, CrossAttnDownBlock2D):
|
1047 |
-
|
1048 |
-
|
1049 |
-
|
1050 |
-
|
1051 |
-
|
|
|
|
|
1052 |
elif isinstance(module, DownBlock2D):
|
1053 |
module.forward = hacked_DownBlock2D_forward.__get__(
|
1054 |
module, DownBlock2D
|
@@ -1057,14 +968,17 @@ class StableDiffusionReferencePipeline:
|
|
1057 |
# module.forward = hacked_CrossAttnUpBlock2D_forward.__get__(module, CrossAttnUpBlock2D)
|
1058 |
# module.mean_bank0 = []
|
1059 |
# module.var_bank0 = []
|
|
|
1060 |
elif isinstance(module, UpBlock2D):
|
1061 |
module.forward = hacked_UpBlock2D_forward.__get__(
|
1062 |
module, UpBlock2D
|
1063 |
)
|
1064 |
module.mean_bank0 = []
|
1065 |
module.var_bank0 = []
|
|
|
1066 |
module.mean_bank = []
|
1067 |
module.var_bank = []
|
|
|
1068 |
module.attention_auto_machine_weight = (
|
1069 |
self.attention_auto_machine_weight
|
1070 |
)
|
@@ -1079,6 +993,7 @@ class StableDiffusionReferencePipeline:
|
|
1079 |
module.style_fidelity = self.style_fidelity
|
1080 |
module.ref_mask = self.ref_mask
|
1081 |
module.inpaint_mask = self.inpaint_mask
|
|
|
1082 |
else:
|
1083 |
gn_modules = None
|
1084 |
elif model_type == "controlnet":
|
@@ -1098,6 +1013,7 @@ class StableDiffusionReferencePipeline:
|
|
1098 |
module, BasicTransformerBlock
|
1099 |
)
|
1100 |
module.bank = []
|
|
|
1101 |
# float(i) / float(len(attn_modules))
|
1102 |
module.attn_weight = 0.0
|
1103 |
module.attention_auto_machine_weight = (
|
@@ -1113,9 +1029,61 @@ class StableDiffusionReferencePipeline:
|
|
1113 |
module.uc_mask = self.uc_mask
|
1114 |
module.style_fidelity = self.style_fidelity
|
1115 |
module.ref_mask = self.ref_mask
|
|
|
1116 |
else:
|
1117 |
attn_modules = None
|
1118 |
-
gn_modules = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1119 |
|
1120 |
return attn_modules, gn_modules
|
1121 |
|
@@ -1123,6 +1091,7 @@ class StableDiffusionReferencePipeline:
|
|
1123 |
if attn_modules is not None:
|
1124 |
for i, module in enumerate(attn_modules):
|
1125 |
module.MODE = mode
|
|
|
1126 |
if gn_modules is not None:
|
1127 |
for i, module in enumerate(gn_modules):
|
1128 |
module.MODE = mode
|
|
|
1 |
# Based on https://raw.githubusercontent.com/okotaku/diffusers/feature/reference_only_control/examples/community/stable_diffusion_reference.py
|
2 |
# Inspired by: https://github.com/Mikubill/sd-webui-controlnet/discussions/1236
|
3 |
+
import torch.fft as fft
|
4 |
from typing import Any, Callable, Dict, List, Optional, Union, Tuple
|
5 |
|
6 |
import numpy as np
|
7 |
import PIL.Image
|
8 |
import torch
|
9 |
|
|
|
10 |
from diffusers.models.attention import BasicTransformerBlock
|
11 |
from diffusers.models.unet_2d_blocks import (
|
12 |
CrossAttnDownBlock2D,
|
|
|
14 |
DownBlock2D,
|
15 |
UpBlock2D,
|
16 |
)
|
|
|
17 |
from diffusers.utils import PIL_INTERPOLATION, logging
|
18 |
import torch.nn.functional as F
|
19 |
|
|
|
20 |
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
21 |
|
22 |
EXAMPLE_DOC_STRING = """
|
|
|
54 |
return result
|
55 |
|
56 |
|
57 |
+
@torch.no_grad()
|
58 |
+
def add_freq_feature(feature1, feature2, ref_ratio):
|
59 |
+
"""
|
60 |
+
feature1: reference feature
|
61 |
+
feature2: target feature
|
62 |
+
ref_ratio: larger ratio means larger reference frequency
|
63 |
+
"""
|
64 |
+
# Convert features to float32 (if not already) for compatibility with fft operations
|
65 |
+
data_type = feature2.dtype
|
66 |
+
feature1 = feature1.to(torch.float32)
|
67 |
+
feature2 = feature2.to(torch.float32)
|
68 |
+
|
69 |
+
# Compute the Fourier transforms of both features
|
70 |
+
spectrum1 = fft.fftn(feature1, dim=(-2, -1))
|
71 |
+
spectrum2 = fft.fftn(feature2, dim=(-2, -1))
|
72 |
+
|
73 |
+
# Extract high-frequency magnitude and phase from feature1
|
74 |
+
magnitude1 = torch.abs(spectrum1)
|
75 |
+
# phase1 = torch.angle(spectrum1)
|
76 |
+
|
77 |
+
# Extract magnitude and phase from feature2
|
78 |
+
magnitude2 = torch.abs(spectrum2)
|
79 |
+
phase2 = torch.angle(spectrum2)
|
80 |
+
|
81 |
+
magnitude2.mul_((1-ref_ratio)).add_(magnitude1 * ref_ratio)
|
82 |
+
# phase2.mul_(1.0).add_(phase1 * 0.0)
|
83 |
+
|
84 |
+
# Combine magnitude and phase information
|
85 |
+
mixed_spectrum = torch.polar(magnitude2, phase2)
|
86 |
+
|
87 |
+
# Compute the inverse Fourier transform to get the mixed feature
|
88 |
+
mixed_feature = fft.ifftn(mixed_spectrum, dim=(-2, -1))
|
89 |
+
|
90 |
+
del feature1, feature2, spectrum1, spectrum2, magnitude1, magnitude2, phase2, mixed_spectrum
|
91 |
+
|
92 |
+
# Convert back to the original data type and return the result
|
93 |
+
return mixed_feature.to(data_type)
|
94 |
+
|
95 |
+
|
96 |
+
@torch.no_grad()
|
97 |
+
def save_ref_feature(feature, mask):
|
98 |
+
"""
|
99 |
+
feature: n,c,h,w
|
100 |
+
mask: n,1,h,w
|
101 |
+
|
102 |
+
return n,c,h,w
|
103 |
+
"""
|
104 |
+
return feature * mask
|
105 |
+
|
106 |
+
|
107 |
+
@torch.no_grad()
|
108 |
+
def mix_ref_feature(feature, ref_fea_bank, cfg=True, ref_scale=0.0, dim3=False):
|
109 |
+
"""
|
110 |
+
feature: n,l,c or n,c,h,w
|
111 |
+
ref_fea_bank: [(n,c,h,w)]
|
112 |
+
cfg: True/False
|
113 |
+
|
114 |
+
return n,l,c or n,c,h,w
|
115 |
+
"""
|
116 |
+
if cfg:
|
117 |
+
ref_fea = torch.cat(
|
118 |
+
(ref_fea_bank+ref_fea_bank), dim=0)
|
119 |
+
else:
|
120 |
+
ref_fea = ref_fea_bank
|
121 |
+
|
122 |
+
if dim3:
|
123 |
+
feature = feature.permute(0, 2, 1).view(ref_fea.shape)
|
124 |
+
|
125 |
+
mixed_feature = add_freq_feature(ref_fea, feature, ref_scale)
|
126 |
+
|
127 |
+
if dim3:
|
128 |
+
mixed_feature = mixed_feature.view(
|
129 |
+
ref_fea.shape[0], ref_fea.shape[1], -1).permute(0, 2, 1)
|
130 |
+
|
131 |
+
del ref_fea
|
132 |
+
del feature
|
133 |
+
return mixed_feature
|
134 |
+
|
135 |
+
|
136 |
+
def mix_norm_feature(x, inpaint_mask, mean_bank, var_bank, do_classifier_free_guidance, style_fidelity, uc_mask, eps=1e-6):
|
137 |
+
"""
|
138 |
+
x: input feature n,c,h,w
|
139 |
+
inpaint_mask: mask region to inpain
|
140 |
+
"""
|
141 |
+
|
142 |
+
# get the inpainting region and only mix this region.
|
143 |
+
scale_ratio = inpaint_mask.shape[2] / x.shape[2]
|
144 |
+
this_inpaint_mask = F.interpolate(
|
145 |
+
inpaint_mask.to(x.device), scale_factor=1 / scale_ratio
|
146 |
+
)
|
147 |
+
this_inpaint_mask = this_inpaint_mask.repeat(
|
148 |
+
x.shape[0], x.shape[1], 1, 1
|
149 |
+
).bool()
|
150 |
+
masked_x = (
|
151 |
+
x[this_inpaint_mask]
|
152 |
+
.detach()
|
153 |
+
.clone()
|
154 |
+
.view(x.shape[0], x.shape[1], -1, 1)
|
155 |
+
)
|
156 |
+
var, mean = torch.var_mean(
|
157 |
+
masked_x, dim=(2, 3), keepdim=True, correction=0
|
158 |
+
)
|
159 |
+
std = torch.maximum(
|
160 |
+
var, torch.zeros_like(var) + eps) ** 0.5
|
161 |
+
mean_acc = sum(mean_bank) / float(len(mean_bank))
|
162 |
+
var_acc = sum(var_bank) / float(len(var_bank))
|
163 |
+
std_acc = (
|
164 |
+
torch.maximum(var_acc, torch.zeros_like(
|
165 |
+
var_acc) + eps) ** 0.5
|
166 |
+
)
|
167 |
+
|
168 |
+
x_uc = (((masked_x - mean) / std) * std_acc) + mean_acc
|
169 |
+
x_c = x_uc.clone()
|
170 |
+
if do_classifier_free_guidance and style_fidelity > 0:
|
171 |
+
x_c[uc_mask] = masked_x[uc_mask]
|
172 |
+
masked_x = style_fidelity * x_c + \
|
173 |
+
(1.0 - style_fidelity) * x_uc
|
174 |
+
x[this_inpaint_mask] = masked_x.view(-1)
|
175 |
+
return x
|
176 |
+
|
177 |
+
|
178 |
class StableDiffusionReferencePipeline:
|
179 |
def prepare_ref_image(
|
180 |
self,
|
|
|
356 |
this_ref_mask = F.interpolate(
|
357 |
this_ref_mask, scale_factor=ref_scale
|
358 |
)
|
359 |
+
self.fea_bank.append(save_ref_feature(
|
360 |
+
resize_norm_hidden_states, this_ref_mask))
|
|
|
361 |
this_ref_mask = this_ref_mask.repeat(
|
362 |
resize_norm_hidden_states.shape[0],
|
363 |
resize_norm_hidden_states.shape[1],
|
|
|
374 |
-1,
|
375 |
)
|
376 |
)
|
377 |
+
|
378 |
masked_norm_hidden_states = masked_norm_hidden_states.permute(
|
379 |
0, 2, 1
|
380 |
)
|
381 |
self.bank.append(masked_norm_hidden_states)
|
382 |
+
del masked_norm_hidden_states
|
383 |
+
del this_ref_mask
|
384 |
+
del resize_norm_hidden_states
|
385 |
attn_output = self.attn1(
|
386 |
norm_hidden_states,
|
387 |
encoder_hidden_states=encoder_hidden_states
|
|
|
392 |
)
|
393 |
if self.MODE == "read":
|
394 |
if self.attention_auto_machine_weight > self.attn_weight:
|
395 |
+
freq_norm_hidden_states = mix_ref_feature(
|
396 |
+
norm_hidden_states,
|
397 |
+
self.fea_bank,
|
398 |
+
cfg=self.do_classifier_free_guidance,
|
399 |
+
ref_scale=self.ref_scale,
|
400 |
+
dim3=True)
|
401 |
+
self.fea_bank.clear()
|
402 |
+
|
403 |
+
this_bank = torch.cat(self.bank+self.bank, dim=0)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
404 |
ref_hidden_states = torch.cat(
|
405 |
+
(freq_norm_hidden_states, this_bank), dim=1
|
406 |
)
|
407 |
+
del this_bank
|
408 |
+
self.bank.clear()
|
409 |
+
|
410 |
attn_output_uc = self.attn1(
|
411 |
+
freq_norm_hidden_states,
|
412 |
encoder_hidden_states=ref_hidden_states,
|
|
|
413 |
**cross_attention_kwargs,
|
414 |
)
|
415 |
+
del ref_hidden_states
|
416 |
attn_output_c = attn_output_uc.clone()
|
417 |
if self.do_classifier_free_guidance and self.style_fidelity > 0:
|
418 |
attn_output_c[self.uc_mask] = self.attn1(
|
|
|
425 |
+ (1.0 - self.style_fidelity) * attn_output_uc
|
426 |
)
|
427 |
self.bank.clear()
|
428 |
+
self.fea_bank.clear()
|
429 |
+
del attn_output_c
|
430 |
+
del attn_output_uc
|
431 |
else:
|
432 |
attn_output = self.attn1(
|
433 |
norm_hidden_states,
|
|
|
437 |
attention_mask=attention_mask,
|
438 |
**cross_attention_kwargs,
|
439 |
)
|
440 |
+
self.bank.clear()
|
441 |
+
self.fea_bank.clear()
|
442 |
+
|
443 |
if self.use_ada_layer_norm_zero:
|
444 |
attn_output = gate_msa.unsqueeze(1) * attn_output
|
445 |
hidden_states = attn_output + hidden_states
|
|
|
488 |
this_ref_mask = F.interpolate(
|
489 |
self.ref_mask.to(x.device), scale_factor=1 / scale_ratio
|
490 |
)
|
491 |
+
|
492 |
+
self.fea_bank.append(save_ref_feature(
|
493 |
+
x, this_ref_mask))
|
494 |
+
|
495 |
this_ref_mask = this_ref_mask.repeat(
|
496 |
x.shape[0], x.shape[1], 1, 1
|
497 |
).bool()
|
|
|
505 |
masked_x, dim=(2, 3), keepdim=True, correction=0
|
506 |
)
|
507 |
|
508 |
+
self.mean_bank.append(torch.cat([mean]*2, dim=0))
|
509 |
+
self.var_bank.append(torch.cat([var]*2, dim=0))
|
510 |
if self.MODE == "read":
|
511 |
if (
|
512 |
self.gn_auto_machine_weight >= self.gn_weight
|
|
|
514 |
and len(self.var_bank) > 0
|
515 |
):
|
516 |
# print("hacked_mid_forward")
|
517 |
+
x = mix_ref_feature(
|
518 |
+
x, self.fea_bank, cfg=self.do_classifier_free_guidance, ref_scale=self.ref_scale)
|
519 |
+
self.fea_bank = []
|
520 |
+
x = mix_norm_feature(x, self.inpaint_mask, self.mean_bank, self.var_bank,
|
521 |
+
self.do_classifier_free_guidance,
|
522 |
+
self.style_fidelity, self.uc_mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
523 |
self.mean_bank = []
|
524 |
self.var_bank = []
|
525 |
return x
|
|
|
550 |
self.ref_mask.to(hidden_states.device),
|
551 |
scale_factor=1 / scale_ratio,
|
552 |
)
|
553 |
+
self.fea_bank0.append(save_ref_feature(
|
554 |
+
hidden_states, this_ref_mask))
|
555 |
this_ref_mask = this_ref_mask.repeat(
|
556 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
557 |
).bool()
|
|
|
564 |
var, mean = torch.var_mean(
|
565 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
566 |
)
|
567 |
+
self.mean_bank0.append(torch.cat([mean]*2, dim=0))
|
568 |
+
self.var_bank0.append(torch.cat([var]*2, dim=0))
|
569 |
if self.MODE == "read":
|
570 |
if (
|
571 |
self.gn_auto_machine_weight >= self.gn_weight
|
|
|
573 |
and len(self.var_bank0) > 0
|
574 |
):
|
575 |
# print("hacked_CrossAttnDownBlock2D_forward0")
|
576 |
+
hidden_states = mix_ref_feature(
|
577 |
+
hidden_states, [self.fea_bank0[i]], cfg=self.do_classifier_free_guidance, ref_scale=self.ref_scale)
|
578 |
+
|
579 |
+
hidden_states = mix_norm_feature(hidden_states, self.inpaint_mask, self.mean_bank0[i], self.var_bank0[i],
|
580 |
+
self.do_classifier_free_guidance,
|
581 |
+
self.style_fidelity, self.uc_mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
582 |
|
583 |
hidden_states = attn(
|
584 |
hidden_states,
|
585 |
encoder_hidden_states=encoder_hidden_states,
|
586 |
cross_attention_kwargs=cross_attention_kwargs,
|
|
|
|
|
587 |
return_dict=False,
|
588 |
)[0]
|
589 |
if self.MODE == "write":
|
|
|
595 |
self.ref_mask.to(hidden_states.device),
|
596 |
scale_factor=1 / scale_ratio,
|
597 |
)
|
598 |
+
self.fea_bank.append(save_ref_feature(
|
599 |
+
hidden_states, this_ref_mask))
|
600 |
this_ref_mask = this_ref_mask.repeat(
|
601 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
602 |
).bool()
|
|
|
609 |
var, mean = torch.var_mean(
|
610 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
611 |
)
|
612 |
+
self.mean_bank.append(torch.cat([mean]*2, dim=0))
|
613 |
+
self.var_bank.append(torch.cat([var]*2, dim=0))
|
614 |
if self.MODE == "read":
|
615 |
if (
|
616 |
self.gn_auto_machine_weight >= self.gn_weight
|
|
|
618 |
and len(self.var_bank) > 0
|
619 |
):
|
620 |
# print("hack_CrossAttnDownBlock2D_forward")
|
621 |
+
hidden_states = mix_ref_feature(
|
622 |
+
hidden_states, [self.fea_bank[i]], cfg=self.do_classifier_free_guidance, ref_scale=self.ref_scale)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
623 |
|
624 |
+
hidden_states = mix_norm_feature(hidden_states, self.inpaint_mask, self.mean_bank[i], self.var_bank[i],
|
625 |
+
self.do_classifier_free_guidance,
|
626 |
+
self.style_fidelity, self.uc_mask)
|
627 |
output_states = output_states + (hidden_states,)
|
628 |
|
629 |
if self.MODE == "read":
|
|
|
631 |
self.var_bank0 = []
|
632 |
self.mean_bank = []
|
633 |
self.var_bank = []
|
634 |
+
self.fea_bank0 = []
|
635 |
+
self.fea_bank = []
|
636 |
|
637 |
if self.downsamplers is not None:
|
638 |
for downsampler in self.downsamplers:
|
|
|
660 |
self.ref_mask.to(hidden_states.device),
|
661 |
scale_factor=1 / scale_ratio,
|
662 |
)
|
663 |
+
self.fea_bank.append(save_ref_feature(
|
664 |
+
hidden_states, this_ref_mask))
|
665 |
this_ref_mask = this_ref_mask.repeat(
|
666 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
667 |
).bool()
|
|
|
674 |
var, mean = torch.var_mean(
|
675 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
676 |
)
|
677 |
+
self.mean_bank.append(torch.cat([mean]*2, dim=0))
|
678 |
+
self.var_bank.append(torch.cat([var]*2, dim=0))
|
679 |
if self.MODE == "read":
|
680 |
if (
|
681 |
self.gn_auto_machine_weight >= self.gn_weight
|
|
|
683 |
and len(self.var_bank) > 0
|
684 |
):
|
685 |
# print("hacked_DownBlock2D_forward")
|
686 |
+
hidden_states = mix_ref_feature(
|
687 |
+
hidden_states, [self.fea_bank[i]], cfg=self.do_classifier_free_guidance, ref_scale=self.ref_scale)
|
688 |
+
|
689 |
+
hidden_states = mix_norm_feature(hidden_states, self.inpaint_mask, self.mean_bank[i], self.var_bank[i],
|
690 |
+
self.do_classifier_free_guidance,
|
691 |
+
self.style_fidelity, self.uc_mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
692 |
|
693 |
output_states = output_states + (hidden_states,)
|
694 |
|
695 |
if self.MODE == "read":
|
696 |
self.mean_bank = []
|
697 |
self.var_bank = []
|
698 |
+
self.fea_bank = []
|
699 |
|
700 |
if self.downsamplers is not None:
|
701 |
for downsampler in self.downsamplers:
|
|
|
736 |
self.ref_mask.to(hidden_states.device),
|
737 |
scale_factor=1 / scale_ratio,
|
738 |
)
|
739 |
+
self.fea_bank0.append(save_ref_feature(
|
740 |
+
hidden_states, this_ref_mask))
|
741 |
this_ref_mask = this_ref_mask.repeat(
|
742 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
743 |
).bool()
|
|
|
750 |
var, mean = torch.var_mean(
|
751 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
752 |
)
|
753 |
+
self.mean_bank0.append(torch.cat([mean]*2, dim=0))
|
754 |
+
self.var_bank0.append(torch.cat([var]*2, dim=0))
|
755 |
if self.MODE == "read":
|
756 |
if (
|
757 |
self.gn_auto_machine_weight >= self.gn_weight
|
|
|
759 |
and len(self.var_bank0) > 0
|
760 |
):
|
761 |
# print("hacked_CrossAttnUpBlock2D_forward1")
|
762 |
+
hidden_states = mix_ref_feature(
|
763 |
+
hidden_states, [self.fea_bank0[i]], cfg=self.do_classifier_free_guidance, ref_scale=self.ref_scale)
|
764 |
+
|
765 |
+
hidden_states = mix_norm_feature(hidden_states, self.inpaint_mask, self.mean_bank0[i], self.var_bank0[i],
|
766 |
+
self.do_classifier_free_guidance,
|
767 |
+
self.style_fidelity, self.uc_mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
768 |
|
769 |
hidden_states = attn(
|
770 |
hidden_states,
|
|
|
785 |
self.ref_mask.to(hidden_states.device),
|
786 |
scale_factor=1 / scale_ratio,
|
787 |
)
|
788 |
+
self.fea_bank.append(save_ref_feature(
|
789 |
+
hidden_states, this_ref_mask))
|
790 |
this_ref_mask = this_ref_mask.repeat(
|
791 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
792 |
).bool()
|
|
|
799 |
var, mean = torch.var_mean(
|
800 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
801 |
)
|
802 |
+
self.mean_bank.append(torch.cat([mean]*2, dim=0))
|
803 |
+
self.var_bank.append(torch.cat([var]*2, dim=0))
|
804 |
if self.MODE == "read":
|
805 |
if (
|
806 |
self.gn_auto_machine_weight >= self.gn_weight
|
|
|
808 |
and len(self.var_bank) > 0
|
809 |
):
|
810 |
# print("hacked_CrossAttnUpBlock2D_forward")
|
811 |
+
hidden_states = mix_ref_feature(
|
812 |
+
hidden_states, [self.fea_bank[i]], cfg=self.do_classifier_free_guidance, ref_scale=self.ref_scale)
|
813 |
+
|
814 |
+
hidden_states = mix_norm_feature(hidden_states, self.inpaint_mask, self.mean_bank[i], self.var_bank[i],
|
815 |
+
self.do_classifier_free_guidance,
|
816 |
+
self.style_fidelity, self.uc_mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
817 |
|
818 |
if self.MODE == "read":
|
819 |
self.mean_bank0 = []
|
820 |
self.var_bank0 = []
|
821 |
self.mean_bank = []
|
822 |
self.var_bank = []
|
823 |
+
self.fea_bank = []
|
824 |
+
self.fea_bank0 = []
|
825 |
|
826 |
if self.upsamplers is not None:
|
827 |
for upsampler in self.upsamplers:
|
|
|
851 |
self.ref_mask.to(hidden_states.device),
|
852 |
scale_factor=1 / scale_ratio,
|
853 |
)
|
854 |
+
self.fea_bank.append(save_ref_feature(
|
855 |
+
hidden_states, this_ref_mask))
|
856 |
this_ref_mask = this_ref_mask.repeat(
|
857 |
hidden_states.shape[0], hidden_states.shape[1], 1, 1
|
858 |
).bool()
|
|
|
865 |
var, mean = torch.var_mean(
|
866 |
masked_hidden_states, dim=(2, 3), keepdim=True, correction=0
|
867 |
)
|
868 |
+
self.mean_bank.append(torch.cat([mean]*2, dim=0))
|
869 |
+
self.var_bank.append(torch.cat([var]*2, dim=0))
|
870 |
if self.MODE == "read":
|
871 |
if (
|
872 |
self.gn_auto_machine_weight >= self.gn_weight
|
|
|
874 |
and len(self.var_bank) > 0
|
875 |
):
|
876 |
# print("hacked_UpBlock2D_forward")
|
877 |
+
hidden_states = mix_ref_feature(
|
878 |
+
hidden_states, [self.fea_bank[i]], cfg=self.do_classifier_free_guidance, ref_scale=self.ref_scale)
|
879 |
+
|
880 |
+
hidden_states = mix_norm_feature(hidden_states, self.inpaint_mask, self.mean_bank[i], self.var_bank[i],
|
881 |
+
self.do_classifier_free_guidance,
|
882 |
+
self.style_fidelity, self.uc_mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
883 |
|
884 |
if self.MODE == "read":
|
885 |
self.mean_bank = []
|
886 |
self.var_bank = []
|
887 |
+
self.fea_bank = []
|
888 |
|
889 |
if self.upsamplers is not None:
|
890 |
for upsampler in self.upsamplers:
|
|
|
910 |
module, BasicTransformerBlock
|
911 |
)
|
912 |
module.bank = []
|
913 |
+
module.fea_bank = []
|
914 |
module.attn_weight = float(i) / float(len(attn_modules))
|
915 |
module.attention_auto_machine_weight = (
|
916 |
self.attention_auto_machine_weight
|
|
|
925 |
module.uc_mask = self.uc_mask
|
926 |
module.style_fidelity = self.style_fidelity
|
927 |
module.ref_mask = self.ref_mask
|
928 |
+
module.ref_scale = self.ref_scale
|
929 |
else:
|
930 |
attn_modules = None
|
931 |
if reference_adain:
|
|
|
952 |
module.forward = hacked_mid_forward.__get__(
|
953 |
module, torch.nn.Module
|
954 |
)
|
955 |
+
# elif isinstance(module, CrossAttnDownBlock2D):
|
956 |
+
# module.forward = hack_CrossAttnDownBlock2D_forward.__get__(
|
957 |
+
# module, CrossAttnDownBlock2D
|
958 |
+
# )
|
959 |
+
# module.mean_bank0 = []
|
960 |
+
# module.var_bank0 = []
|
961 |
+
# module.fea_bank0 = []
|
962 |
+
|
963 |
elif isinstance(module, DownBlock2D):
|
964 |
module.forward = hacked_DownBlock2D_forward.__get__(
|
965 |
module, DownBlock2D
|
|
|
968 |
# module.forward = hacked_CrossAttnUpBlock2D_forward.__get__(module, CrossAttnUpBlock2D)
|
969 |
# module.mean_bank0 = []
|
970 |
# module.var_bank0 = []
|
971 |
+
# module.fea_bank0 = []
|
972 |
elif isinstance(module, UpBlock2D):
|
973 |
module.forward = hacked_UpBlock2D_forward.__get__(
|
974 |
module, UpBlock2D
|
975 |
)
|
976 |
module.mean_bank0 = []
|
977 |
module.var_bank0 = []
|
978 |
+
module.fea_bank0 = []
|
979 |
module.mean_bank = []
|
980 |
module.var_bank = []
|
981 |
+
module.fea_bank = []
|
982 |
module.attention_auto_machine_weight = (
|
983 |
self.attention_auto_machine_weight
|
984 |
)
|
|
|
993 |
module.style_fidelity = self.style_fidelity
|
994 |
module.ref_mask = self.ref_mask
|
995 |
module.inpaint_mask = self.inpaint_mask
|
996 |
+
module.ref_scale = self.ref_scale
|
997 |
else:
|
998 |
gn_modules = None
|
999 |
elif model_type == "controlnet":
|
|
|
1013 |
module, BasicTransformerBlock
|
1014 |
)
|
1015 |
module.bank = []
|
1016 |
+
module.fea_bank = []
|
1017 |
# float(i) / float(len(attn_modules))
|
1018 |
module.attn_weight = 0.0
|
1019 |
module.attention_auto_machine_weight = (
|
|
|
1029 |
module.uc_mask = self.uc_mask
|
1030 |
module.style_fidelity = self.style_fidelity
|
1031 |
module.ref_mask = self.ref_mask
|
1032 |
+
module.ref_scale = self.ref_scale
|
1033 |
else:
|
1034 |
attn_modules = None
|
1035 |
+
# gn_modules = None
|
1036 |
+
if reference_adain:
|
1037 |
+
gn_modules = [model.mid_block]
|
1038 |
+
model.mid_block.gn_weight = 0
|
1039 |
+
|
1040 |
+
down_blocks = model.down_blocks
|
1041 |
+
for w, module in enumerate(down_blocks):
|
1042 |
+
module.gn_weight = 1.0 - float(w) / float(len(down_blocks))
|
1043 |
+
gn_modules.append(module)
|
1044 |
+
# print(module.__class__.__name__,module.gn_weight)
|
1045 |
+
|
1046 |
+
|
1047 |
+
for i, module in enumerate(gn_modules):
|
1048 |
+
if getattr(module, "original_forward", None) is None:
|
1049 |
+
module.original_forward = module.forward
|
1050 |
+
if i == 0:
|
1051 |
+
# mid_block
|
1052 |
+
module.forward = hacked_mid_forward.__get__(
|
1053 |
+
module, torch.nn.Module
|
1054 |
+
)
|
1055 |
+
# elif isinstance(module, CrossAttnDownBlock2D):
|
1056 |
+
# module.forward = hack_CrossAttnDownBlock2D_forward.__get__(
|
1057 |
+
# module, CrossAttnDownBlock2D
|
1058 |
+
# )
|
1059 |
+
# module.mean_bank0 = []
|
1060 |
+
# module.var_bank0 = []
|
1061 |
+
# module.fea_bank0 = []
|
1062 |
+
|
1063 |
+
elif isinstance(module, DownBlock2D):
|
1064 |
+
module.forward = hacked_DownBlock2D_forward.__get__(
|
1065 |
+
module, DownBlock2D
|
1066 |
+
)
|
1067 |
+
module.mean_bank = []
|
1068 |
+
module.var_bank = []
|
1069 |
+
module.fea_bank = []
|
1070 |
+
module.attention_auto_machine_weight = (
|
1071 |
+
self.attention_auto_machine_weight
|
1072 |
+
)
|
1073 |
+
module.gn_auto_machine_weight = self.gn_auto_machine_weight
|
1074 |
+
module.do_classifier_free_guidance = (
|
1075 |
+
self.do_classifier_free_guidance
|
1076 |
+
)
|
1077 |
+
module.do_classifier_free_guidance = (
|
1078 |
+
self.do_classifier_free_guidance
|
1079 |
+
)
|
1080 |
+
module.uc_mask = self.uc_mask
|
1081 |
+
module.style_fidelity = self.style_fidelity
|
1082 |
+
module.ref_mask = self.ref_mask
|
1083 |
+
module.inpaint_mask = self.inpaint_mask
|
1084 |
+
module.ref_scale = self.ref_scale
|
1085 |
+
else:
|
1086 |
+
gn_modules = None
|
1087 |
|
1088 |
return attn_modules, gn_modules
|
1089 |
|
|
|
1091 |
if attn_modules is not None:
|
1092 |
for i, module in enumerate(attn_modules):
|
1093 |
module.MODE = mode
|
1094 |
+
|
1095 |
if gn_modules is not None:
|
1096 |
for i, module in enumerate(gn_modules):
|
1097 |
module.MODE = mode
|