Upload two_shot.py
Browse files- two_shot.py +227 -0
two_shot.py
ADDED
@@ -0,0 +1,227 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Dict, Optional, Tuple
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
|
6 |
+
from modules import devices
|
7 |
+
|
8 |
+
import modules.scripts as scripts
|
9 |
+
import gradio as gr
|
10 |
+
# todo:
|
11 |
+
from modules.script_callbacks import CFGDenoisedParams, on_cfg_denoised
|
12 |
+
|
13 |
+
from modules.processing import StableDiffusionProcessing
|
14 |
+
|
15 |
+
|
16 |
+
@dataclass
|
17 |
+
class Division:
|
18 |
+
y: float
|
19 |
+
x: float
|
20 |
+
|
21 |
+
|
22 |
+
@dataclass
|
23 |
+
class Position:
|
24 |
+
y: float
|
25 |
+
x: float
|
26 |
+
ey: float
|
27 |
+
ex: float
|
28 |
+
|
29 |
+
|
30 |
+
class Filter:
|
31 |
+
|
32 |
+
def __init__(self, division: Division, position: Position, weight: float):
|
33 |
+
self.division = division
|
34 |
+
self.position = position
|
35 |
+
self.weight = weight
|
36 |
+
|
37 |
+
def create_tensor(self, num_channels: int, height_b: int, width_b: int) -> torch.Tensor:
|
38 |
+
|
39 |
+
x = torch.zeros(num_channels, height_b, width_b).to(devices.device)
|
40 |
+
|
41 |
+
division_height = height_b / self.division.y
|
42 |
+
division_width = width_b / self.division.x
|
43 |
+
y1 = int(division_height * self.position.y)
|
44 |
+
y2 = int(division_height * self.position.ey)
|
45 |
+
x1 = int(division_width * self.position.x)
|
46 |
+
x2 = int(division_width * self.position.ex)
|
47 |
+
|
48 |
+
x[:, y1:y2, x1:x2] = self.weight
|
49 |
+
|
50 |
+
return x
|
51 |
+
|
52 |
+
|
53 |
+
class Script(scripts.Script):
|
54 |
+
|
55 |
+
def __init__(self):
|
56 |
+
self.num_batches: int = 0
|
57 |
+
self.end_at_step: int = 20
|
58 |
+
self.filters: List[Filter] = []
|
59 |
+
self.debug: bool = False
|
60 |
+
|
61 |
+
def title(self):
|
62 |
+
return "Latent Couple extension"
|
63 |
+
|
64 |
+
def show(self, is_img2img):
|
65 |
+
return scripts.AlwaysVisible
|
66 |
+
|
67 |
+
def create_filters_from_ui_params(self, raw_divisions: str, raw_positions: str, raw_weights: str):
|
68 |
+
|
69 |
+
divisions = []
|
70 |
+
for division in raw_divisions.split(','):
|
71 |
+
y, x = division.split(':')
|
72 |
+
divisions.append(Division(float(y), float(x)))
|
73 |
+
|
74 |
+
def start_and_end_position(raw: str):
|
75 |
+
nums = [float(num) for num in raw.split('-')]
|
76 |
+
if len(nums) == 1:
|
77 |
+
return nums[0], nums[0] + 1.0
|
78 |
+
else:
|
79 |
+
return nums[0], nums[1]
|
80 |
+
|
81 |
+
positions = []
|
82 |
+
for position in raw_positions.split(','):
|
83 |
+
y, x = position.split(':')
|
84 |
+
y1, y2 = start_and_end_position(y)
|
85 |
+
x1, x2 = start_and_end_position(x)
|
86 |
+
positions.append(Position(y1, x1, y2, x2))
|
87 |
+
|
88 |
+
weights = []
|
89 |
+
for w in raw_weights.split(','):
|
90 |
+
weights.append(float(w))
|
91 |
+
|
92 |
+
# todo: assert len
|
93 |
+
|
94 |
+
return [Filter(division, position, weight) for division, position, weight in zip(divisions, positions, weights)]
|
95 |
+
|
96 |
+
def do_visualize(self, raw_divisions: str, raw_positions: str, raw_weights: str):
|
97 |
+
|
98 |
+
self.filters = self.create_filters_from_ui_params(raw_divisions, raw_positions, raw_weights)
|
99 |
+
|
100 |
+
return [f.create_tensor(1, 128, 128).squeeze(dim=0).cpu().numpy() for f in self.filters]
|
101 |
+
|
102 |
+
def do_apply(self, extra_generation_params: str):
|
103 |
+
#
|
104 |
+
# parse "Latent Couple" extra_generation_params
|
105 |
+
#
|
106 |
+
raw_params = {}
|
107 |
+
|
108 |
+
for assignment in extra_generation_params.split(' '):
|
109 |
+
pair = assignment.split('=', 1)
|
110 |
+
if len(pair) != 2:
|
111 |
+
continue
|
112 |
+
raw_params[pair[0]] = pair[1]
|
113 |
+
|
114 |
+
return raw_params.get('divisions', '1:1,1:2,1:2'), raw_params.get('positions', '0:0,0:0,0:1'), raw_params.get('weights', '0.2,0.8,0.8'), int(raw_params.get('step', '20'))
|
115 |
+
|
116 |
+
def ui(self, is_img2img):
|
117 |
+
id_part = "img2img" if is_img2img else "txt2img"
|
118 |
+
|
119 |
+
with gr.Group():
|
120 |
+
with gr.Accordion("Latent Couple", open=False):
|
121 |
+
enabled = gr.Checkbox(value=False, label="Enabled")
|
122 |
+
with gr.Row():
|
123 |
+
divisions = gr.Textbox(label="Divisions", elem_id=f"cd_{id_part}_divisions", value="1:1,1:2,1:2")
|
124 |
+
positions = gr.Textbox(label="Positions", elem_id=f"cd_{id_part}_positions", value="0:0,0:0,0:1")
|
125 |
+
with gr.Row():
|
126 |
+
weights = gr.Textbox(label="Weights", elem_id=f"cd_{id_part}_weights", value="0.2,0.8,0.8")
|
127 |
+
end_at_step = gr.Slider(minimum=0, maximum=150, step=1, label="end at this step", elem_id=f"cd_{id_part}_end_at_this_step", value=20)
|
128 |
+
|
129 |
+
visualize_button = gr.Button(value="Visualize")
|
130 |
+
visual_regions = gr.Gallery(label="Regions").style(grid=(4, 4, 4, 8), height="auto")
|
131 |
+
|
132 |
+
visualize_button.click(fn=self.do_visualize, inputs=[divisions, positions, weights], outputs=[visual_regions])
|
133 |
+
|
134 |
+
extra_generation_params = gr.Textbox(label="Extra generation params")
|
135 |
+
apply_button = gr.Button(value="Apply")
|
136 |
+
|
137 |
+
apply_button.click(fn=self.do_apply, inputs=[extra_generation_params], outputs=[divisions, positions, weights, end_at_step])
|
138 |
+
|
139 |
+
self.infotext_fields = [
|
140 |
+
(extra_generation_params, "Latent Couple")
|
141 |
+
]
|
142 |
+
return enabled, divisions, positions, weights, end_at_step
|
143 |
+
|
144 |
+
def denoised_callback(self, params: CFGDenoisedParams):
|
145 |
+
|
146 |
+
if self.enabled and params.sampling_step < self.end_at_step:
|
147 |
+
|
148 |
+
x = params.x
|
149 |
+
# x.shape = [batch_size, C, H // 8, W // 8]
|
150 |
+
|
151 |
+
num_batches = self.num_batches
|
152 |
+
num_prompts = x.shape[0] // num_batches
|
153 |
+
# ex. num_batches = 3
|
154 |
+
# ex. num_prompts = 3 (tensor) + 1 (uncond)
|
155 |
+
|
156 |
+
if self.debug:
|
157 |
+
print(f"### Latent couple ###")
|
158 |
+
print(f"denoised_callback x.shape={x.shape} num_batches={num_batches} num_prompts={num_prompts}")
|
159 |
+
|
160 |
+
filters = [
|
161 |
+
f.create_tensor(x.shape[1], x.shape[2], x.shape[3]) for f in self.filters
|
162 |
+
]
|
163 |
+
neg_filters = [1.0 - f for f in filters]
|
164 |
+
|
165 |
+
"""
|
166 |
+
batch #1
|
167 |
+
subprompt #1
|
168 |
+
subprompt #2
|
169 |
+
subprompt #3
|
170 |
+
batch #2
|
171 |
+
subprompt #1
|
172 |
+
subprompt #2
|
173 |
+
subprompt #3
|
174 |
+
uncond
|
175 |
+
batch #1
|
176 |
+
batch #2
|
177 |
+
"""
|
178 |
+
|
179 |
+
tensor_off = 0
|
180 |
+
uncond_off = num_batches * num_prompts - num_batches
|
181 |
+
for b in range(num_batches):
|
182 |
+
uncond = x[uncond_off, :, :, :]
|
183 |
+
|
184 |
+
for p in range(num_prompts - 1):
|
185 |
+
if self.debug:
|
186 |
+
print(f"b={b} p={p}")
|
187 |
+
if p < len(filters):
|
188 |
+
tensor = x[tensor_off, :, :, :]
|
189 |
+
x[tensor_off, :, :, :] = tensor * filters[p] + uncond * neg_filters[p]
|
190 |
+
|
191 |
+
tensor_off += 1
|
192 |
+
|
193 |
+
uncond_off += 1
|
194 |
+
|
195 |
+
def process(self, p: StableDiffusionProcessing, enabled: bool, raw_divisions: str, raw_positions: str, raw_weights: str, raw_end_at_step: int):
|
196 |
+
|
197 |
+
self.enabled = enabled
|
198 |
+
|
199 |
+
if not self.enabled:
|
200 |
+
return
|
201 |
+
|
202 |
+
self.num_batches = p.batch_size
|
203 |
+
|
204 |
+
self.filters = self.create_filters_from_ui_params(raw_divisions, raw_positions, raw_weights)
|
205 |
+
|
206 |
+
self.end_at_step = raw_end_at_step
|
207 |
+
|
208 |
+
#
|
209 |
+
|
210 |
+
if self.end_at_step != 0:
|
211 |
+
p.extra_generation_params["Latent Couple"] = f"divisions={raw_divisions} positions={raw_positions} weights={raw_weights} end at step={raw_end_at_step}"
|
212 |
+
# save params into the output file as PNG textual data.
|
213 |
+
|
214 |
+
if self.debug:
|
215 |
+
print(f"### Latent couple ###")
|
216 |
+
print(f"process num_batches={self.num_batches} end_at_step={self.end_at_step}")
|
217 |
+
|
218 |
+
if not hasattr(self, 'callbacks_added'):
|
219 |
+
on_cfg_denoised(self.denoised_callback)
|
220 |
+
self.callbacks_added = True
|
221 |
+
|
222 |
+
return
|
223 |
+
|
224 |
+
def postprocess(self, *args):
|
225 |
+
return
|
226 |
+
|
227 |
+
|