Sergio19997 commited on
Commit
2933ada
Β·
verified Β·
1 Parent(s): 34942c8

Upload nodes.py

Browse files
Files changed (1) hide show
  1. nodes.py +1701 -0
nodes.py ADDED
@@ -0,0 +1,1701 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os, glob, sys
2
+ import logging
3
+
4
+ import torch
5
+ import torch.nn.functional as F
6
+ import torchvision.transforms as T
7
+ from torchvision.transforms.functional import normalize
8
+ from torchvision.ops import masks_to_boxes
9
+
10
+ import numpy as np
11
+ import cv2
12
+ import math
13
+ from typing import List
14
+ from PIL import Image
15
+ import io
16
+ from scipy import stats
17
+ from insightface.app.common import Face
18
+ from segment_anything import sam_model_registry
19
+
20
+ from modules.processing import ProcessingImg2Img
21
+ from modules.shared import state
22
+ # from comfy_extras.chainner_models import model_loading
23
+ import comfy.model_management as model_management
24
+ import comfy.utils
25
+ import folder_paths
26
+
27
+ import scripts.reactor_version
28
+ from r_chainner import model_loading
29
+ from scripts.reactor_faceswap import (
30
+ FaceSwapScript,
31
+ get_models,
32
+ get_current_faces_model,
33
+ analyze_faces,
34
+ half_det_size,
35
+ providers
36
+ )
37
+ from scripts.reactor_swapper import (
38
+ unload_all_models,
39
+ )
40
+ from scripts.reactor_logger import logger
41
+ from reactor_utils import (
42
+ batch_tensor_to_pil,
43
+ batched_pil_to_tensor,
44
+ tensor_to_pil,
45
+ img2tensor,
46
+ tensor2img,
47
+ save_face_model,
48
+ load_face_model,
49
+ download,
50
+ set_ort_session,
51
+ prepare_cropped_face,
52
+ normalize_cropped_face,
53
+ add_folder_path_and_extensions,
54
+ rgba2rgb_tensor,
55
+ progress_bar,
56
+ progress_bar_reset
57
+ )
58
+ from reactor_patcher import apply_patch
59
+ from r_facelib.utils.face_restoration_helper import FaceRestoreHelper
60
+ from r_basicsr.utils.registry import ARCH_REGISTRY
61
+ import scripts.r_archs.codeformer_arch
62
+ import scripts.r_masking.subcore as subcore
63
+ import scripts.r_masking.core as core
64
+ import scripts.r_masking.segs as masking_segs
65
+
66
+ import scripts.reactor_sfw as sfw
67
+
68
+
69
+ models_dir = folder_paths.models_dir
70
+ REACTOR_MODELS_PATH = os.path.join(models_dir, "reactor")
71
+ FACE_MODELS_PATH = os.path.join(REACTOR_MODELS_PATH, "faces")
72
+ NSFWDET_MODEL_PATH = os.path.join(models_dir, "nsfw_detector","vit-base-nsfw-detector")
73
+
74
+ if not os.path.exists(REACTOR_MODELS_PATH):
75
+ os.makedirs(REACTOR_MODELS_PATH)
76
+ if not os.path.exists(FACE_MODELS_PATH):
77
+ os.makedirs(FACE_MODELS_PATH)
78
+
79
+ dir_facerestore_models = os.path.join(models_dir, "facerestore_models")
80
+ os.makedirs(dir_facerestore_models, exist_ok=True)
81
+ folder_paths.folder_names_and_paths["facerestore_models"] = ([dir_facerestore_models], folder_paths.supported_pt_extensions)
82
+
83
+ BLENDED_FACE_MODEL = None
84
+ FACE_SIZE: int = 512
85
+ FACE_HELPER = None
86
+
87
+ if "ultralytics" not in folder_paths.folder_names_and_paths:
88
+ add_folder_path_and_extensions("ultralytics_bbox", [os.path.join(models_dir, "ultralytics", "bbox")], folder_paths.supported_pt_extensions)
89
+ add_folder_path_and_extensions("ultralytics_segm", [os.path.join(models_dir, "ultralytics", "segm")], folder_paths.supported_pt_extensions)
90
+ add_folder_path_and_extensions("ultralytics", [os.path.join(models_dir, "ultralytics")], folder_paths.supported_pt_extensions)
91
+ if "sams" not in folder_paths.folder_names_and_paths:
92
+ add_folder_path_and_extensions("sams", [os.path.join(models_dir, "sams")], folder_paths.supported_pt_extensions)
93
+
94
+ def get_facemodels():
95
+ models_path = os.path.join(FACE_MODELS_PATH, "*")
96
+ models = glob.glob(models_path)
97
+ models = [x for x in models if x.endswith(".safetensors")]
98
+ return models
99
+
100
+ def get_restorers():
101
+ models_path = os.path.join(models_dir, "facerestore_models/*")
102
+ models = glob.glob(models_path)
103
+ models = [x for x in models if (x.endswith(".pth") or x.endswith(".onnx"))]
104
+ if len(models) == 0:
105
+ fr_urls = [
106
+ "https://huggingface.co/datasets/Gourieff/ReActor/resolve/main/models/facerestore_models/GFPGANv1.3.pth",
107
+ "https://huggingface.co/datasets/Gourieff/ReActor/resolve/main/models/facerestore_models/GFPGANv1.4.pth",
108
+ "https://huggingface.co/datasets/Gourieff/ReActor/resolve/main/models/facerestore_models/codeformer-v0.1.0.pth",
109
+ "https://huggingface.co/datasets/Gourieff/ReActor/resolve/main/models/facerestore_models/GPEN-BFR-512.onnx",
110
+ ]
111
+ for model_url in fr_urls:
112
+ model_name = os.path.basename(model_url)
113
+ model_path = os.path.join(dir_facerestore_models, model_name)
114
+ download(model_url, model_path, model_name)
115
+ models = glob.glob(models_path)
116
+ models = [x for x in models if (x.endswith(".pth") or x.endswith(".onnx"))]
117
+ return models
118
+
119
+ def get_model_names(get_models):
120
+ models = get_models()
121
+ names = []
122
+ for x in models:
123
+ names.append(os.path.basename(x))
124
+ names.sort(key=str.lower)
125
+ names.insert(0, "none")
126
+ return names
127
+
128
+ def model_names():
129
+ models = get_models()
130
+ return {os.path.basename(x): x for x in models}
131
+
132
+
133
+ class reactor:
134
+ @classmethod
135
+ def INPUT_TYPES(s):
136
+ return {
137
+ "required": {
138
+ "enabled": ("BOOLEAN", {"default": True, "label_off": "OFF", "label_on": "ON"}),
139
+ "input_image": ("IMAGE",),
140
+ "swap_model": (list(model_names().keys()),),
141
+ "facedetection": (["retinaface_resnet50", "retinaface_mobile0.25", "YOLOv5l", "YOLOv5n"],),
142
+ "face_restore_model": (get_model_names(get_restorers),),
143
+ "face_restore_visibility": ("FLOAT", {"default": 1, "min": 0.1, "max": 1, "step": 0.05}),
144
+ "codeformer_weight": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1, "step": 0.05}),
145
+ "detect_gender_input": (["no","female","male"], {"default": "no"}),
146
+ "detect_gender_source": (["no","female","male"], {"default": "no"}),
147
+ "input_faces_index": ("STRING", {"default": "0"}),
148
+ "source_faces_index": ("STRING", {"default": "0"}),
149
+ "console_log_level": ([0, 1, 2], {"default": 1}),
150
+ },
151
+ "optional": {
152
+ "source_image": ("IMAGE",),
153
+ "face_model": ("FACE_MODEL",),
154
+ "face_boost": ("FACE_BOOST",),
155
+ },
156
+ "hidden": {"faces_order": "FACES_ORDER"},
157
+ }
158
+
159
+ RETURN_TYPES = ("IMAGE","FACE_MODEL","IMAGE")
160
+ RETURN_NAMES = ("SWAPPED_IMAGE","FACE_MODEL","ORIGINAL_IMAGE")
161
+ FUNCTION = "execute"
162
+ CATEGORY = "🌌 ReActor"
163
+
164
+ def __init__(self):
165
+ # self.face_helper = None
166
+ self.faces_order = ["large-small", "large-small"]
167
+ # self.face_size = FACE_SIZE
168
+ self.face_boost_enabled = False
169
+ self.restore = True
170
+ self.boost_model = None
171
+ self.interpolation = "Bicubic"
172
+ self.boost_model_visibility = 1
173
+ self.boost_cf_weight = 0.5
174
+ self.last_swapped_bboxes = None
175
+ # self.last_swapped_indices = None
176
+ self.restore_swapped_only = True
177
+
178
+ def restore_face(
179
+ self,
180
+ input_image,
181
+ face_restore_model,
182
+ face_restore_visibility,
183
+ codeformer_weight,
184
+ facedetection,
185
+ ):
186
+
187
+ # from datetime import datetime
188
+ # def _current_time():
189
+ # current_datetime = datetime.now()
190
+ # return current_datetime.strftime("%M:%S")
191
+
192
+ # локальная функция IoU
193
+ def _iou(b1, b2):
194
+ xA = max(b1[0], b2[0])
195
+ yA = max(b1[1], b2[1])
196
+ xB = min(b1[2], b2[2])
197
+ yB = min(b1[3], b2[3])
198
+ interW = max(0.0, xB - xA)
199
+ interH = max(0.0, yB - yA)
200
+ interArea = interW * interH
201
+ if interArea == 0:
202
+ return 0.0
203
+ boxAArea = max(1.0, (b1[2]-b1[0]) * (b1[3]-b1[1]))
204
+ boxBArea = max(1.0, (b2[2]-b2[0]) * (b2[3]-b2[1]))
205
+ return interArea / float(boxAArea + boxBArea - interArea)
206
+
207
+ result = input_image
208
+
209
+ if face_restore_model != "none" and not model_management.processing_interrupted():
210
+
211
+ global FACE_SIZE, FACE_HELPER
212
+
213
+ self.face_helper = FACE_HELPER
214
+
215
+ faceSize = 512
216
+ if "1024" in face_restore_model.lower():
217
+ faceSize = 1024
218
+ elif "2048" in face_restore_model.lower():
219
+ faceSize = 2048
220
+
221
+ logger.status(f"Restoring with {face_restore_model} | Face Size is set to {faceSize}")
222
+
223
+ model_path = folder_paths.get_full_path("facerestore_models", face_restore_model)
224
+
225
+ device = model_management.get_torch_device()
226
+
227
+ if "codeformer" in face_restore_model.lower():
228
+
229
+ codeformer_net = ARCH_REGISTRY.get("CodeFormer")(
230
+ dim_embd=512,
231
+ codebook_size=1024,
232
+ n_head=8,
233
+ n_layers=9,
234
+ connect_list=["32", "64", "128", "256"],
235
+ ).to(device)
236
+ checkpoint = torch.load(model_path)["params_ema"]
237
+ codeformer_net.load_state_dict(checkpoint)
238
+ facerestore_model = codeformer_net.eval()
239
+
240
+ elif ".onnx" in face_restore_model:
241
+
242
+ ort_session = set_ort_session(model_path, providers=providers)
243
+ ort_session_inputs = {}
244
+ facerestore_model = ort_session
245
+
246
+ else:
247
+
248
+ sd = comfy.utils.load_torch_file(model_path, safe_load=True)
249
+ facerestore_model = model_loading.load_state_dict(sd).eval()
250
+ facerestore_model.to(device)
251
+
252
+ if faceSize != FACE_SIZE or self.face_helper is None:
253
+ self.face_helper = FaceRestoreHelper(1, face_size=faceSize, crop_ratio=(1, 1), det_model=facedetection, save_ext='png', use_parse=True, device=device)
254
+ FACE_SIZE = faceSize
255
+ FACE_HELPER = self.face_helper
256
+
257
+ # Copying Tensor to CPU (if it isn't) to convert torch.Tensor to np.ndarray
258
+ image_np = 255. * result.cpu().numpy()
259
+
260
+ total_images = image_np.shape[0]
261
+
262
+ out_images = []
263
+
264
+ pbar = progress_bar(total_images)
265
+
266
+ for i in range(total_images):
267
+
268
+ # if total_images > 1:
269
+ # logger.status(f"Restoring {i}")
270
+
271
+ cur_image_np = image_np[i,:, :, ::-1]
272
+
273
+ original_resolution = cur_image_np.shape[0:2]
274
+
275
+ if facerestore_model is None or self.face_helper is None:
276
+ return result
277
+
278
+ self.face_helper.clean_all()
279
+ self.face_helper.read_image(cur_image_np)
280
+ self.face_helper.get_face_landmarks_5(only_center_face=False, resize=640, eye_dist_threshold=5)
281
+ self.face_helper.align_warp_face()
282
+
283
+ # restored_face = None
284
+ restored_faces = []
285
+
286
+ # Π±Π΅Ρ€Π΅ΠΌ сохранённыС bbox ΠΈΠ· swap (ΠΈΠ»ΠΈ None)
287
+ swapped_bboxes = getattr(self, "last_swapped_bboxes", None)
288
+ # Ρ„Π»Π°Π³ΠΈ, Ρ‡Ρ‚ΠΎΠ±Ρ‹ ΠΎΠ΄Π½ΠΎ сохранённоС bbox Π½Π΅ совпало с нСсколькими Π»ΠΈΡ†Π°ΠΌΠΈ
289
+ used_swapped = [False] * len(swapped_bboxes) if swapped_bboxes else None
290
+
291
+ IOU_THRESHOLD = 0.5
292
+
293
+ for idx, cropped_face in enumerate(self.face_helper.cropped_faces):
294
+
295
+ # опрСдСляСм bbox Ρ‚Π΅ΠΊΡƒΡ‰Π΅Π³ΠΎ Π»ΠΈΡ†Π°, ΠΊΠΎΡ‚ΠΎΡ€Ρ‹ΠΉ Π΄Π°Π» Π΄Π΅Ρ‚Π΅ΠΊΡ‚ΠΎΡ€ Π²Π½ΡƒΡ‚Ρ€ΠΈ FaceRestoreHelper
296
+ current_bbox = None
297
+ if hasattr(self.face_helper, 'det_faces') and len(self.face_helper.det_faces) > idx:
298
+ det = self.face_helper.det_faces[idx]
299
+ # det ΠΌ.Π±. [x1,y1,x2,y2,score]
300
+ current_bbox = (float(det[0]), float(det[1]), float(det[2]), float(det[3]))
301
+
302
+ # Π»ΠΎΠ³ΠΈΠΊΠ°: Ссли Ρƒ нас Π΅ΡΡ‚ΡŒ сохранённыС bbox β€” рСсторим Ρ‚ΠΎΠ»ΡŒΠΊΠΎ Ссли iou с ΠΎΠ΄Π½ΠΈΠΌ ΠΈΠ· Π½ΠΈΡ… > ΠΏΠΎΡ€ΠΎΠ³
303
+ do_restore = True
304
+ # if self.last_swapped_indices is not None:
305
+ # do_restore = idx in self.last_swapped_indices
306
+ if swapped_bboxes and self.restore_swapped_only:
307
+ do_restore = False
308
+ if current_bbox is not None:
309
+ for s_idx, sbox in enumerate(swapped_bboxes):
310
+ if used_swapped is not None and used_swapped[s_idx]:
311
+ continue
312
+ if _iou(current_bbox, sbox) >= IOU_THRESHOLD:
313
+ cx1 = (current_bbox[0] + current_bbox[2]) / 2
314
+ cy1 = (current_bbox[1] + current_bbox[3]) / 2
315
+ cx2 = (sbox[0] + sbox[2]) / 2
316
+ cy2 = (sbox[1] + sbox[3]) / 2
317
+ if abs(cx1 - cx2) < (current_bbox[2] - current_bbox[0]) * 0.25:
318
+ if abs(cy1 - cy2) < (current_bbox[3] - current_bbox[1]) * 0.25:
319
+ do_restore = True
320
+ # do_restore = True
321
+ if used_swapped is not None:
322
+ used_swapped[s_idx] = True
323
+ break
324
+
325
+ if do_restore:
326
+
327
+ # if ".pth" in face_restore_model:
328
+ cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True)
329
+ normalize(cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
330
+ cropped_face_t = cropped_face_t.unsqueeze(0).to(device)
331
+
332
+ try:
333
+
334
+ with torch.no_grad():
335
+
336
+ if ".onnx" in face_restore_model: # ONNX models
337
+
338
+ for ort_session_input in ort_session.get_inputs():
339
+ if ort_session_input.name == "input":
340
+ cropped_face_prep = prepare_cropped_face(cropped_face)
341
+ ort_session_inputs[ort_session_input.name] = cropped_face_prep
342
+ if ort_session_input.name == "weight":
343
+ weight = np.array([ 1 ], dtype = np.double)
344
+ ort_session_inputs[ort_session_input.name] = weight
345
+
346
+ output = ort_session.run(None, ort_session_inputs)[0][0]
347
+ restored_face = normalize_cropped_face(output)
348
+
349
+ else: # PTH models
350
+
351
+ output = facerestore_model(cropped_face_t, w=codeformer_weight)[0] if "codeformer" in face_restore_model.lower() else facerestore_model(cropped_face_t)[0]
352
+ restored_face = tensor2img(output, rgb2bgr=True, min_max=(-1, 1))
353
+
354
+ del output
355
+ torch.cuda.empty_cache()
356
+
357
+ except Exception as error:
358
+
359
+ print(f"\tFailed inference: {error}", file=sys.stderr)
360
+ # restored_face = tensor2img(cropped_face_t, rgb2bgr=True, min_max=(-1, 1))
361
+ restored_face = cropped_face.copy()
362
+
363
+ else:
364
+ restored_face = cropped_face.copy()
365
+
366
+ if face_restore_visibility < 1:
367
+ restored_face = cropped_face * (1 - face_restore_visibility) + restored_face * face_restore_visibility
368
+
369
+ restored_face = restored_face.astype("uint8")
370
+ self.face_helper.add_restored_face(restored_face)
371
+
372
+ self.face_helper.get_inverse_affine(None)
373
+
374
+ restored_img = self.face_helper.paste_faces_to_input_image()
375
+ restored_img = restored_img[:, :, ::-1]
376
+
377
+ if original_resolution != restored_img.shape[0:2]:
378
+ restored_img = cv2.resize(restored_img, (0, 0), fx=original_resolution[1]/restored_img.shape[1], fy=original_resolution[0]/restored_img.shape[0], interpolation=cv2.INTER_AREA)
379
+
380
+ self.face_helper.clean_all()
381
+
382
+ # out_images[i] = restored_img
383
+ out_images.append(restored_img)
384
+
385
+ if state.interrupted or model_management.processing_interrupted():
386
+ logger.status("Interrupted by User")
387
+ return input_image
388
+
389
+ pbar.update(1)
390
+
391
+ restored_img_np = np.array(out_images).astype(np.float32) / 255.0
392
+ restored_img_tensor = torch.from_numpy(restored_img_np)
393
+
394
+ result = restored_img_tensor
395
+
396
+ progress_bar_reset(pbar)
397
+
398
+ # if hasattr(self, "last_swapped_bboxes"):
399
+ self.last_swapped_bboxes = None
400
+ # if hasattr(self, "last_swapped_indices"):
401
+ # self.last_swapped_indices = None
402
+
403
+ return result
404
+
405
+
406
+ def execute(self, enabled, input_image, swap_model, detect_gender_source, detect_gender_input, source_faces_index, input_faces_index, console_log_level, face_restore_model,face_restore_visibility, codeformer_weight, facedetection, source_image=None, face_model=None, faces_order=None, face_boost=None):
407
+
408
+ device = model_management.get_torch_device()
409
+
410
+ if isinstance(input_image, torch.Tensor) and input_image.device != device:
411
+ input_image = input_image.to(device)
412
+
413
+ if face_boost is not None:
414
+ self.face_boost_enabled = face_boost["enabled"]
415
+ self.boost_model = face_boost["boost_model"]
416
+ self.interpolation = face_boost["interpolation"]
417
+ self.boost_model_visibility = face_boost["visibility"]
418
+ self.boost_cf_weight = face_boost["codeformer_weight"]
419
+ self.restore = face_boost["restore_with_main_after"]
420
+ else:
421
+ self.face_boost_enabled = False
422
+
423
+ if faces_order is None:
424
+ faces_order = self.faces_order
425
+
426
+ apply_patch(console_log_level)
427
+
428
+ if not enabled:
429
+ return (input_image,face_model)
430
+ elif source_image is None and face_model is None:
431
+ logger.error("Please provide 'source_image' or `face_model`")
432
+ return (input_image,face_model)
433
+
434
+ if face_model == "none":
435
+ face_model = None
436
+
437
+ script = FaceSwapScript()
438
+ pil_images = batch_tensor_to_pil(input_image)
439
+
440
+
441
+ logger.status("Checking for any unsafe content...")
442
+ pbar = progress_bar(len(pil_images))
443
+ pil_images_sfw = []
444
+ for img in pil_images:
445
+ if state.interrupted or model_management.processing_interrupted():
446
+ logger.status("Interrupted by User")
447
+ break
448
+ img_byte_arr = io.BytesIO()
449
+ img.save(img_byte_arr, format='PNG')
450
+ img_byte_arr = img_byte_arr.getvalue()
451
+ if not sfw.nsfw_image(img_byte_arr, NSFWDET_MODEL_PATH):
452
+ pil_images_sfw.append(img)
453
+ pbar.update(1)
454
+ pil_images = pil_images_sfw
455
+
456
+ progress_bar_reset(pbar)
457
+
458
+ if len(pil_images) > 0:
459
+
460
+ if source_image is not None:
461
+ source = tensor_to_pil(source_image)
462
+ else:
463
+ source = None
464
+
465
+ p = ProcessingImg2Img(pil_images)
466
+ script.process(
467
+ p=p,
468
+ img=source,
469
+ enable=True,
470
+ source_faces_index=source_faces_index,
471
+ faces_index=input_faces_index,
472
+ model=swap_model,
473
+ swap_in_source=True,
474
+ swap_in_generated=True,
475
+ gender_source=detect_gender_source,
476
+ gender_target=detect_gender_input,
477
+ face_model=face_model,
478
+ faces_order=faces_order,
479
+ # face boost:
480
+ face_boost_enabled=self.face_boost_enabled,
481
+ face_restore_model=self.boost_model,
482
+ face_restore_visibility=self.boost_model_visibility,
483
+ codeformer_weight=self.boost_cf_weight,
484
+ interpolation=self.interpolation,
485
+ )
486
+ result = batched_pil_to_tensor(p.init_images)
487
+ # print(f"bbox={p.bbox}")
488
+ if len(p.bbox) > 0:
489
+ self.last_swapped_bboxes = p.bbox
490
+ # self.last_swapped_indices = p.swapped_indexes
491
+ original_image = input_image
492
+
493
+ if face_model is None:
494
+ current_face_model = get_current_faces_model()
495
+ face_model_to_provide = current_face_model[0] if (current_face_model is not None and len(current_face_model) > 0) else face_model
496
+ else:
497
+ face_model_to_provide = face_model
498
+
499
+ if self.restore or not self.face_boost_enabled:
500
+ result = reactor.restore_face(self,result,face_restore_model,face_restore_visibility,codeformer_weight,facedetection)
501
+
502
+ else:
503
+ image_black = Image.new("RGB", (512, 512))
504
+ result = batched_pil_to_tensor([image_black])
505
+ face_model_to_provide = None
506
+ original_image = result
507
+
508
+ return (result,face_model_to_provide,original_image)
509
+
510
+
511
+ class ReActorPlusOpt:
512
+ @classmethod
513
+ def INPUT_TYPES(s):
514
+ return {
515
+ "required": {
516
+ "enabled": ("BOOLEAN", {"default": True, "label_off": "OFF", "label_on": "ON"}),
517
+ "input_image": ("IMAGE",),
518
+ "swap_model": (list(model_names().keys()),),
519
+ "facedetection": (["retinaface_resnet50", "retinaface_mobile0.25", "YOLOv5l", "YOLOv5n"],),
520
+ "face_restore_model": (get_model_names(get_restorers),),
521
+ "face_restore_visibility": ("FLOAT", {"default": 1, "min": 0.1, "max": 1, "step": 0.05}),
522
+ "codeformer_weight": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1, "step": 0.05}),
523
+ },
524
+ "optional": {
525
+ "source_image": ("IMAGE",),
526
+ "face_model": ("FACE_MODEL",),
527
+ "options": ("OPTIONS",),
528
+ "face_boost": ("FACE_BOOST",),
529
+ }
530
+ }
531
+
532
+ RETURN_TYPES = ("IMAGE","FACE_MODEL","IMAGE")
533
+ RETURN_NAMES = ("SWAPPED_IMAGE","FACE_MODEL","ORIGINAL_IMAGE")
534
+ FUNCTION = "execute"
535
+ CATEGORY = "🌌 ReActor"
536
+
537
+ def __init__(self):
538
+ # self.face_helper = None
539
+ self.faces_order = ["large-small", "large-small"]
540
+ self.detect_gender_input = "no"
541
+ self.detect_gender_source = "no"
542
+ self.input_faces_index = "0"
543
+ self.source_faces_index = "0"
544
+ self.console_log_level = 1
545
+ self.restore_swapped_only = True
546
+ # self.face_size = 512
547
+ self.face_boost_enabled = False
548
+ self.restore = True
549
+ self.boost_model = None
550
+ self.interpolation = "Bicubic"
551
+ self.boost_model_visibility = 1
552
+ self.boost_cf_weight = 0.5
553
+
554
+ def execute(self, enabled, input_image, swap_model, facedetection, face_restore_model, face_restore_visibility, codeformer_weight, source_image=None, face_model=None, options=None, face_boost=None):
555
+
556
+ if options is not None:
557
+ self.faces_order = [options["input_faces_order"], options["source_faces_order"]]
558
+ self.console_log_level = options["console_log_level"]
559
+ self.detect_gender_input = options["detect_gender_input"]
560
+ self.detect_gender_source = options["detect_gender_source"]
561
+ self.input_faces_index = options["input_faces_index"]
562
+ self.source_faces_index = options["source_faces_index"]
563
+ self.restore_swapped_only = options["restore_swapped_only"]
564
+
565
+ if face_boost is not None:
566
+ self.face_boost_enabled = face_boost["enabled"]
567
+ self.restore = face_boost["restore_with_main_after"]
568
+ else:
569
+ self.face_boost_enabled = False
570
+
571
+ result = reactor.execute(
572
+ self,enabled,input_image,swap_model,self.detect_gender_source,self.detect_gender_input,self.source_faces_index,self.input_faces_index,self.console_log_level,face_restore_model,face_restore_visibility,codeformer_weight,facedetection,source_image,face_model,self.faces_order, face_boost=face_boost
573
+ )
574
+
575
+ return result
576
+
577
+
578
+ class LoadFaceModel:
579
+ @classmethod
580
+ def INPUT_TYPES(s):
581
+ return {
582
+ "required": {
583
+ "face_model": (get_model_names(get_facemodels),),
584
+ }
585
+ }
586
+
587
+ RETURN_TYPES = ("FACE_MODEL","STRING")
588
+ RETURN_NAMES = ("FACE_MODEL","FACE_MODEL_NAME")
589
+ FUNCTION = "load_model"
590
+ CATEGORY = "🌌 ReActor"
591
+
592
+ def load_model(self, face_model):
593
+ self.face_model = face_model
594
+ face_model = face_model.split(".safetensors")[0] if ".safetensors" in face_model else face_model
595
+ self.face_models_path = FACE_MODELS_PATH
596
+ if self.face_model != "none":
597
+ face_model_path = os.path.join(self.face_models_path, self.face_model)
598
+ out = load_face_model(face_model_path)
599
+ else:
600
+ out = None
601
+ return (out,face_model)
602
+
603
+
604
+ class ReActorWeight:
605
+ @classmethod
606
+ def INPUT_TYPES(s):
607
+ return {
608
+ "required": {
609
+ "input_image": ("IMAGE",),
610
+ "faceswap_weight": (["0%", "12.5%", "25%", "37.5%", "50%", "62.5%", "75%", "87.5%", "100%"], {"default": "50%"}),
611
+ },
612
+ "optional": {
613
+ "source_image": ("IMAGE",),
614
+ "face_model": ("FACE_MODEL",),
615
+ }
616
+ }
617
+
618
+ RETURN_TYPES = ("IMAGE","FACE_MODEL")
619
+ RETURN_NAMES = ("INPUT_IMAGE","FACE_MODEL")
620
+ FUNCTION = "set_weight"
621
+
622
+ OUTPUT_NODE = True
623
+
624
+ CATEGORY = "🌌 ReActor"
625
+
626
+ def set_weight(self, input_image, faceswap_weight, face_model=None, source_image=None):
627
+
628
+ if input_image is None:
629
+ logger.error("Please provide `input_image`")
630
+ return (input_image,None)
631
+
632
+ if source_image is None and face_model is None:
633
+ logger.error("Please provide `source_image` or `face_model`")
634
+ return (input_image,None)
635
+
636
+ weight = float(faceswap_weight.split("%")[0])
637
+
638
+ images = []
639
+ faces = [] if face_model is None else [face_model]
640
+ embeddings = [] if face_model is None else [face_model.embedding]
641
+
642
+ if weight == 0:
643
+ images = [input_image]
644
+ faces = []
645
+ embeddings = []
646
+ elif weight == 100:
647
+ if face_model is None:
648
+ images = [source_image]
649
+ else:
650
+ if weight > 50:
651
+ images = [input_image]
652
+ count = round(100/(100-weight))
653
+ else:
654
+ if face_model is None:
655
+ images = [source_image]
656
+ count = round(100/(weight))
657
+ for i in range(count-1):
658
+ if weight > 50:
659
+ if face_model is None:
660
+ images.append(source_image)
661
+ else:
662
+ faces.append(face_model)
663
+ embeddings.append(face_model.embedding)
664
+ else:
665
+ images.append(input_image)
666
+
667
+ images_list: List[Image.Image] = []
668
+
669
+ apply_patch(1)
670
+
671
+ if len(images) > 0:
672
+
673
+ for image in images:
674
+ img = tensor_to_pil(image)
675
+ images_list.append(img)
676
+
677
+ for image in images_list:
678
+ face = BuildFaceModel.build_face_model(self,image)
679
+ if isinstance(face, str):
680
+ continue
681
+ faces.append(face)
682
+ embeddings.append(face.embedding)
683
+
684
+ if len(faces) > 0:
685
+ blended_embedding = np.mean(embeddings, axis=0)
686
+ blended_face = Face(
687
+ bbox=faces[0].bbox,
688
+ kps=faces[0].kps,
689
+ det_score=faces[0].det_score,
690
+ landmark_3d_68=faces[0].landmark_3d_68,
691
+ pose=faces[0].pose,
692
+ landmark_2d_106=faces[0].landmark_2d_106,
693
+ embedding=blended_embedding,
694
+ gender=faces[0].gender,
695
+ age=faces[0].age
696
+ )
697
+ if blended_face is None:
698
+ no_face_msg = "Something went wrong, please try another set of images"
699
+ logger.error(no_face_msg)
700
+
701
+ return (input_image,blended_face)
702
+
703
+
704
+ class BuildFaceModel:
705
+ def __init__(self):
706
+ self.output_dir = FACE_MODELS_PATH
707
+
708
+ @classmethod
709
+ def INPUT_TYPES(s):
710
+ return {
711
+ "required": {
712
+ "save_mode": ("BOOLEAN", {"default": True, "label_off": "OFF", "label_on": "ON"}),
713
+ "send_only": ("BOOLEAN", {"default": False, "label_off": "NO", "label_on": "YES"}),
714
+ "face_model_name": ("STRING", {"default": "default"}),
715
+ "compute_method": (["Mean", "Median", "Mode"], {"default": "Mean"}),
716
+ },
717
+ "optional": {
718
+ "images": ("IMAGE",),
719
+ "face_models": ("FACE_MODEL",),
720
+ }
721
+ }
722
+
723
+ RETURN_TYPES = ("FACE_MODEL",)
724
+ FUNCTION = "blend_faces"
725
+
726
+ OUTPUT_NODE = True
727
+
728
+ CATEGORY = "🌌 ReActor"
729
+
730
+ def build_face_model(self, image: Image.Image, det_size=(640, 640)):
731
+ logging.StreamHandler.terminator = "\n"
732
+ if image is None:
733
+ error_msg = "Please load an Image"
734
+ logger.error(error_msg)
735
+ return error_msg
736
+ image = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
737
+ face_model = analyze_faces(image, det_size)
738
+
739
+ if len(face_model) == 0:
740
+ print("")
741
+ det_size_half = half_det_size(det_size)
742
+ face_model = analyze_faces(image, det_size_half)
743
+ if face_model is not None and len(face_model) > 0:
744
+ print("...........................................................", end=" ")
745
+
746
+ if face_model is not None and len(face_model) > 0:
747
+ return face_model[0]
748
+ else:
749
+ no_face_msg = "No face found, please try another image"
750
+ # logger.error(no_face_msg)
751
+ return no_face_msg
752
+
753
+ def blend_faces(self, save_mode, send_only, face_model_name, compute_method, images=None, face_models=None):
754
+ global BLENDED_FACE_MODEL
755
+ blended_face: Face = BLENDED_FACE_MODEL
756
+
757
+ if send_only and blended_face is None:
758
+ send_only = False
759
+
760
+ if (images is not None or face_models is not None) and not send_only:
761
+
762
+ faces = []
763
+ embeddings = []
764
+
765
+ apply_patch(1)
766
+
767
+ if images is not None:
768
+ images_list: List[Image.Image] = batch_tensor_to_pil(images)
769
+
770
+ n = len(images_list)
771
+
772
+ for i,image in enumerate(images_list):
773
+ logging.StreamHandler.terminator = " "
774
+ logger.status(f"Building Face Model {i+1} of {n}...")
775
+ face = self.build_face_model(image)
776
+ if isinstance(face, str):
777
+ logger.error(f"No faces found in image {i+1}, skipping")
778
+ continue
779
+ else:
780
+ print(f"{int(((i+1)/n)*100)}%")
781
+ faces.append(face)
782
+ embeddings.append(face.embedding)
783
+
784
+ elif face_models is not None:
785
+
786
+ n = len(face_models)
787
+
788
+ for i,face_model in enumerate(face_models):
789
+ logging.StreamHandler.terminator = " "
790
+ logger.status(f"Extracting Face Model {i+1} of {n}...")
791
+ face = face_model
792
+ if isinstance(face, str):
793
+ logger.error(f"No faces found for face_model {i+1}, skipping")
794
+ continue
795
+ else:
796
+ print(f"{int(((i+1)/n)*100)}%")
797
+ faces.append(face)
798
+ embeddings.append(face.embedding)
799
+
800
+ logging.StreamHandler.terminator = "\n"
801
+ if len(faces) > 0:
802
+ # compute_method_name = "Mean" if compute_method == 0 else "Median" if compute_method == 1 else "Mode"
803
+ logger.status(f"Blending with Compute Method '{compute_method}'...")
804
+ blended_embedding = np.mean(embeddings, axis=0) if compute_method == "Mean" else np.median(embeddings, axis=0) if compute_method == "Median" else stats.mode(embeddings, axis=0)[0].astype(np.float32)
805
+ blended_face = Face(
806
+ bbox=faces[0].bbox,
807
+ kps=faces[0].kps,
808
+ det_score=faces[0].det_score,
809
+ landmark_3d_68=faces[0].landmark_3d_68,
810
+ pose=faces[0].pose,
811
+ landmark_2d_106=faces[0].landmark_2d_106,
812
+ embedding=blended_embedding,
813
+ gender=faces[0].gender,
814
+ age=faces[0].age
815
+ )
816
+ if blended_face is not None:
817
+ BLENDED_FACE_MODEL = blended_face
818
+ if save_mode:
819
+ face_model_path = os.path.join(FACE_MODELS_PATH, face_model_name + ".safetensors")
820
+ save_face_model(blended_face,face_model_path)
821
+ # done_msg = f"Face model has been saved to '{face_model_path}'"
822
+ # logger.status(done_msg)
823
+ logger.status("--Done!--")
824
+ # return (blended_face,)
825
+ else:
826
+ no_face_msg = "Something went wrong, please try another set of images"
827
+ logger.error(no_face_msg)
828
+ # return (blended_face,)
829
+ # logger.status("--Done!--")
830
+ if images is None and face_models is None:
831
+ logger.error("Please provide `images` or `face_models`")
832
+ return (blended_face,)
833
+
834
+
835
+ class SaveFaceModel:
836
+ def __init__(self):
837
+ self.output_dir = FACE_MODELS_PATH
838
+
839
+ @classmethod
840
+ def INPUT_TYPES(s):
841
+ return {
842
+ "required": {
843
+ "save_mode": ("BOOLEAN", {"default": True, "label_off": "OFF", "label_on": "ON"}),
844
+ "face_model_name": ("STRING", {"default": "default"}),
845
+ "select_face_index": ("INT", {"default": 0, "min": 0}),
846
+ },
847
+ "optional": {
848
+ "image": ("IMAGE",),
849
+ "face_model": ("FACE_MODEL",),
850
+ }
851
+ }
852
+
853
+ RETURN_TYPES = ()
854
+ FUNCTION = "save_model"
855
+
856
+ OUTPUT_NODE = True
857
+
858
+ CATEGORY = "🌌 ReActor"
859
+
860
+ def save_model(self, save_mode, face_model_name, select_face_index, image=None, face_model=None, det_size=(640, 640)):
861
+ if save_mode and image is not None:
862
+ source = tensor_to_pil(image)
863
+ source = cv2.cvtColor(np.array(source), cv2.COLOR_RGB2BGR)
864
+ apply_patch(1)
865
+ logger.status("Building Face Model...")
866
+ face_model_raw = analyze_faces(source, det_size)
867
+ if len(face_model_raw) == 0:
868
+ det_size_half = half_det_size(det_size)
869
+ face_model_raw = analyze_faces(source, det_size_half)
870
+ try:
871
+ face_model = face_model_raw[select_face_index]
872
+ except:
873
+ logger.error("No face(s) found")
874
+ return face_model_name
875
+ logger.status("--Done!--")
876
+ if save_mode and (face_model != "none" or face_model is not None):
877
+ face_model_path = os.path.join(self.output_dir, face_model_name + ".safetensors")
878
+ save_face_model(face_model,face_model_path)
879
+ if image is None and face_model is None:
880
+ logger.error("Please provide `face_model` or `image`")
881
+ return face_model_name
882
+
883
+
884
+ class RestoreFace:
885
+ @classmethod
886
+ def INPUT_TYPES(s):
887
+ return {
888
+ "required": {
889
+ "image": ("IMAGE",),
890
+ "facedetection": (["retinaface_resnet50", "retinaface_mobile0.25", "YOLOv5l", "YOLOv5n"],),
891
+ "model": (get_model_names(get_restorers),),
892
+ "visibility": ("FLOAT", {"default": 1, "min": 0.0, "max": 1, "step": 0.05}),
893
+ "codeformer_weight": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1, "step": 0.05}),
894
+ },
895
+ }
896
+
897
+ RETURN_TYPES = ("IMAGE",)
898
+ FUNCTION = "execute"
899
+ CATEGORY = "🌌 ReActor"
900
+
901
+ def execute(self, image, model, visibility, codeformer_weight, facedetection):
902
+ result = reactor.restore_face(
903
+ self, image, model, visibility, codeformer_weight, facedetection
904
+ )
905
+ return (result,)
906
+
907
+
908
+ class RestoreFaceAdvanced:
909
+ @classmethod
910
+ def INPUT_TYPES(s):
911
+ return {
912
+ "required": {
913
+ "image": ("IMAGE",),
914
+ "facedetection": (["retinaface_resnet50", "retinaface_mobile0.25", "YOLOv5l", "YOLOv5n"],),
915
+ "model": (get_model_names(get_restorers),),
916
+ "visibility": ("FLOAT", {"default": 1, "min": 0.0, "max": 1, "step": 0.05}),
917
+ "codeformer_weight": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1, "step": 0.05}),
918
+ "face_selection": (["all", "filter", "largest"],{"default": "all"}),
919
+ },
920
+ "optional": {
921
+ "sort_by": (["area", "x_position", "y_position", "detection_confidence"],{"default": "area"}),
922
+ "reverse_order": ("BOOLEAN", {"default": False}),
923
+ "take_start": ("INT", {"default": 0, "min": 0, "max": 100, "step": 1}),
924
+ "take_count": ("INT", {"default": 1, "min": 1, "max": 100, "step": 1}),
925
+ }
926
+ }
927
+
928
+ RETURN_TYPES = ("IMAGE",)
929
+ FUNCTION = "execute"
930
+ CATEGORY = "🌌 ReActor"
931
+
932
+ def execute(
933
+ self, image, model, visibility, codeformer_weight, facedetection, face_selection, sort_by="area", reverse_order=False, take_start=0, take_count=1
934
+ ):
935
+
936
+ min_x_position=0.0
937
+ max_x_position=1.0
938
+ min_y_position=0.0
939
+ max_y_position=1.0
940
+
941
+ result = image
942
+
943
+ face_restore_model = model
944
+
945
+ if face_restore_model != "none" and not model_management.processing_interrupted():
946
+
947
+ global FACE_SIZE, FACE_HELPER
948
+
949
+ self.face_helper = FACE_HELPER
950
+
951
+ faceSize = 512
952
+ if "1024" in face_restore_model.lower():
953
+ faceSize = 1024
954
+ elif "2048" in face_restore_model.lower():
955
+ faceSize = 2048
956
+
957
+ logger.status(f"Restoring with {face_restore_model} | Face Size is set to {faceSize}")
958
+
959
+ model_path = folder_paths.get_full_path("facerestore_models", face_restore_model)
960
+
961
+ device = model_management.get_torch_device()
962
+
963
+ if "codeformer" in face_restore_model.lower():
964
+
965
+ codeformer_net = ARCH_REGISTRY.get("CodeFormer")(
966
+ dim_embd=512,
967
+ codebook_size=1024,
968
+ n_head=8,
969
+ n_layers=9,
970
+ connect_list=["32", "64", "128", "256"],
971
+ ).to(device)
972
+ checkpoint = torch.load(model_path)["params_ema"]
973
+ codeformer_net.load_state_dict(checkpoint)
974
+ facerestore_model = codeformer_net.eval()
975
+
976
+ elif ".onnx" in face_restore_model:
977
+
978
+ ort_session = set_ort_session(model_path, providers=providers)
979
+ ort_session_inputs = {}
980
+ facerestore_model = ort_session
981
+
982
+ else:
983
+
984
+ sd = comfy.utils.load_torch_file(model_path, safe_load=True)
985
+ facerestore_model = model_loading.load_state_dict(sd).eval()
986
+ facerestore_model.to(device)
987
+
988
+ if faceSize != FACE_SIZE or self.face_helper is None:
989
+ self.face_helper = FaceRestoreHelper(1, face_size=faceSize, crop_ratio=(1, 1), det_model=facedetection, save_ext='png', use_parse=True, device=device)
990
+ FACE_SIZE = faceSize
991
+ FACE_HELPER = self.face_helper
992
+
993
+ image_np = 255. * result.cpu().numpy()
994
+
995
+ total_images = image_np.shape[0]
996
+
997
+ out_images = []
998
+
999
+ pbar = progress_bar(total_images)
1000
+
1001
+ for i in range(total_images):
1002
+
1003
+ cur_image_np = image_np[i,:, :, ::-1]
1004
+
1005
+ original_resolution = cur_image_np.shape[0:2]
1006
+
1007
+ if facerestore_model is None or self.face_helper is None:
1008
+ return result
1009
+
1010
+ self.face_helper.clean_all()
1011
+ self.face_helper.read_image(cur_image_np)
1012
+ self.face_helper.get_face_landmarks_5(only_center_face=False, resize=640, eye_dist_threshold=5)
1013
+ self.face_helper.align_warp_face()
1014
+
1015
+ # Face-Filter Mode
1016
+
1017
+ # Π€ΠΈΠ»ΡŒΡ‚Ρ€Π°Ρ†ΠΈΡ Π»ΠΈΡ†
1018
+ if face_selection != "all" and self.face_helper.cropped_faces:
1019
+ # Π‘ΠΎΠ±ΠΈΡ€Π°Π΅ΠΌ ΠΈΠ½Ρ„ΠΎΡ€ΠΌΠ°Ρ†ΠΈΡŽ ΠΎ Π»ΠΈΡ†Π°Ρ… для Ρ„ΠΈΠ»ΡŒΡ‚Ρ€Π°Ρ†ΠΈΠΈ
1020
+ face_info = []
1021
+ img_height, img_width = cur_image_np.shape[0:2]
1022
+
1023
+ for j, face in enumerate(self.face_helper.cropped_faces):
1024
+ # Π˜ΡΠΏΠΎΠ»ΡŒΠ·ΡƒΠ΅ΠΌ Ρ†Π΅Π½Ρ‚Ρ€ Π»ΠΈΡ†Π° вмСсто Π»Π΅Π²ΠΎΠ³ΠΎ Π²Π΅Ρ€Ρ…Π½Π΅Π³ΠΎ ΡƒΠ³Π»Π°
1025
+ if hasattr(self.face_helper, 'det_faces') and len(self.face_helper.det_faces) > j:
1026
+ bbox = self.face_helper.det_faces[j]
1027
+ # ВычисляСм Ρ†Π΅Π½Ρ‚Ρ€ Π»ΠΈΡ†Π° для Π±ΠΎΠ»Π΅Π΅ Ρ‚ΠΎΡ‡Π½ΠΎΠ³ΠΎ позиционирования
1028
+ x1 = ((bbox[0] + bbox[2]) / 2) / img_width # Ρ†Π΅Π½Ρ‚Ρ€ x
1029
+ y1 = ((bbox[1] + bbox[3]) / 2) / img_height # Ρ†Π΅Π½Ρ‚Ρ€ y
1030
+ area = face.shape[0] * face.shape[1]
1031
+ confidence = bbox[4] if len(bbox) > 4 else 1.0
1032
+ else:
1033
+ # Если информация ΠΎ bbox нСдоступна, ΠΈΡΠΏΠΎΠ»ΡŒΠ·ΡƒΠ΅ΠΌ ΠΏΡ€ΠΈΠ±Π»ΠΈΠ·ΠΈΡ‚Π΅Π»ΡŒΠ½Ρ‹Π΅ Π΄Π°Π½Π½Ρ‹Π΅
1034
+ area = face.shape[0] * face.shape[1]
1035
+ x1, y1 = 0.5, 0.5 # Ρ†Π΅Π½Ρ‚Ρ€ изобраТСния
1036
+ confidence = 1.0
1037
+
1038
+ face_info.append({
1039
+ 'index': j,
1040
+ 'area': area,
1041
+ 'x_position': x1,
1042
+ 'y_position': y1,
1043
+ 'detection_confidence': confidence
1044
+ })
1045
+
1046
+ # Π‘Π½Π°Ρ‡Π°Π»Π° сортируСм всС Π»ΠΈΡ†Π° ΠΏΠΎ Π²Ρ‹Π±Ρ€Π°Π½Π½ΠΎΠΌΡƒ ΠΊΡ€ΠΈΡ‚Π΅Ρ€ΠΈΡŽ
1047
+ all_indices = list(range(len(self.face_helper.cropped_faces)))
1048
+
1049
+ # Π’Ρ‹Π²ΠΎΠ΄ для x_position ΠΈ y_position
1050
+ if sort_by == "y_position":
1051
+ all_positions = [(idx, face_info[idx]['y_position']) for idx in all_indices]
1052
+ elif sort_by == "x_position":
1053
+ all_positions = [(idx, face_info[idx]['x_position']) for idx in all_indices]
1054
+
1055
+ # Π‘ΠΎΡ€Ρ‚ΠΈΡ€ΠΎΠ²ΠΊΠ° ΠΏΠΎ Π²Ρ‹Π±Ρ€Π°Π½Π½ΠΎΠΌΡƒ ΠΊΡ€ΠΈΡ‚Π΅Ρ€ΠΈΡŽ
1056
+ sorted_indices = sorted(
1057
+ all_indices,
1058
+ key=lambda idx: face_info[idx][sort_by],
1059
+ reverse=reverse_order
1060
+ )
1061
+
1062
+ # ΠžΡ‚Π»Π°Π΄ΠΎΡ‡Π½Ρ‹ΠΉ Π²Ρ‹Π²ΠΎΠ΄ послС сортировки
1063
+ if sort_by == "y_position":
1064
+ sorted_positions = [(idx, face_info[idx]['y_position']) for idx in sorted_indices]
1065
+ elif sort_by == "x_position":
1066
+ sorted_positions = [(idx, face_info[idx]['x_position']) for idx in sorted_indices]
1067
+
1068
+ # ΠŸΡ€ΠΈΠΌΠ΅Π½ΡΠ΅ΠΌ Ρ„ΠΈΠ»ΡŒΡ‚Ρ€Π°Ρ†ΠΈΡŽ Π² зависимости ΠΎΡ‚ Ρ€Π΅ΠΆΠΈΠΌΠ°
1069
+ if face_selection == "filter":
1070
+ # Π€ΠΈΠ»ΡŒΡ‚Ρ€Π°Ρ†ΠΈΡ ΠΏΠΎ ΠΊΠΎΠΎΡ€Π΄ΠΈΠ½Π°Ρ‚Π°ΠΌ
1071
+ filtered_indices = [
1072
+ idx for idx in sorted_indices
1073
+ if min_x_position <= face_info[idx]['x_position'] <= max_x_position and
1074
+ min_y_position <= face_info[idx]['y_position'] <= max_y_position
1075
+ ]
1076
+
1077
+ # Π’Ρ‹Π±ΠΎΡ€ΠΊΠ° ΠΏΠΎ take_start ΠΈ take_count
1078
+ selected_indices = filtered_indices[take_start:take_start + take_count]
1079
+
1080
+ elif face_selection == "largest":
1081
+ # ΠŸΡ€ΠΈ Π²Ρ‹Π±ΠΎΡ€Π΅ "largest" просто Π±Π΅Ρ€Π΅ΠΌ take_count Π»ΠΈΡ† с наибольшСй ΠΏΠ»ΠΎΡ‰Π°Π΄ΡŒΡŽ, начиная с take_start
1082
+ selected_indices = sorted_indices[take_start:take_start + take_count]
1083
+
1084
+ elif face_selection == "index":
1085
+ # Π’ Ρ€Π΅ΠΆΠΈΠΌΠ΅ "index" просто Π±Π΅Ρ€Π΅ΠΌ Π»ΠΈΡ†Π°, начиная с take_start
1086
+ selected_indices = sorted_indices[take_start:take_start + take_count]
1087
+
1088
+ if selected_indices:
1089
+ self.face_helper.cropped_faces = [self.face_helper.cropped_faces[j] for j in selected_indices]
1090
+ if hasattr(self.face_helper, 'restored_faces'):
1091
+ self.face_helper.restored_faces = []
1092
+ if hasattr(self.face_helper, 'affine_matrices'):
1093
+ self.face_helper.affine_matrices = [self.face_helper.affine_matrices[j] for j in selected_indices]
1094
+ if hasattr(self.face_helper, 'det_faces'):
1095
+ self.face_helper.det_faces = [self.face_helper.det_faces[j] for j in selected_indices]
1096
+
1097
+ # Face-Filter Mode END
1098
+
1099
+ restored_face = None
1100
+
1101
+ for idx, cropped_face in enumerate(self.face_helper.cropped_faces):
1102
+
1103
+ cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True)
1104
+ normalize(cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
1105
+ cropped_face_t = cropped_face_t.unsqueeze(0).to(device)
1106
+
1107
+ try:
1108
+
1109
+ with torch.no_grad():
1110
+
1111
+ if ".onnx" in face_restore_model: # ONNX models
1112
+
1113
+ for ort_session_input in ort_session.get_inputs():
1114
+ if ort_session_input.name == "input":
1115
+ cropped_face_prep = prepare_cropped_face(cropped_face)
1116
+ ort_session_inputs[ort_session_input.name] = cropped_face_prep
1117
+ if ort_session_input.name == "weight":
1118
+ weight = np.array([ 1 ], dtype = np.double)
1119
+ ort_session_inputs[ort_session_input.name] = weight
1120
+
1121
+ output = ort_session.run(None, ort_session_inputs)[0][0]
1122
+ restored_face = normalize_cropped_face(output)
1123
+
1124
+ else: # PTH models
1125
+
1126
+ output = facerestore_model(cropped_face_t, w=codeformer_weight)[0] if "codeformer" in face_restore_model.lower() else facerestore_model(cropped_face_t)[0]
1127
+ restored_face = tensor2img(output, rgb2bgr=True, min_max=(-1, 1))
1128
+
1129
+ del output
1130
+ torch.cuda.empty_cache()
1131
+
1132
+ except Exception as error:
1133
+
1134
+ print(f"\tFailed inference: {error}", file=sys.stderr)
1135
+ restored_face = cropped_face.copy()
1136
+
1137
+ if visibility < 1:
1138
+ restored_face = cropped_face * (1 - visibility) + restored_face * visibility
1139
+
1140
+ restored_face = restored_face.astype("uint8")
1141
+ self.face_helper.add_restored_face(restored_face)
1142
+
1143
+ self.face_helper.get_inverse_affine(None)
1144
+
1145
+ restored_img = self.face_helper.paste_faces_to_input_image()
1146
+ restored_img = restored_img[:, :, ::-1]
1147
+
1148
+ if original_resolution != restored_img.shape[0:2]:
1149
+ restored_img = cv2.resize(restored_img, (0, 0), fx=original_resolution[1]/restored_img.shape[1], fy=original_resolution[0]/restored_img.shape[0], interpolation=cv2.INTER_AREA)
1150
+
1151
+ self.face_helper.clean_all()
1152
+
1153
+ out_images.append(restored_img)
1154
+
1155
+ if state.interrupted or model_management.processing_interrupted():
1156
+ logger.status("Interrupted by User")
1157
+ return image
1158
+
1159
+ pbar.update(1)
1160
+
1161
+ restored_img_np = np.array(out_images).astype(np.float32) / 255.0
1162
+ restored_img_tensor = torch.from_numpy(restored_img_np)
1163
+
1164
+ result = restored_img_tensor
1165
+
1166
+ progress_bar_reset(pbar)
1167
+
1168
+ return (result,)
1169
+
1170
+
1171
+ class MaskHelper:
1172
+ def __init__(self):
1173
+ self.labels = "all"
1174
+ self.detailer_hook = None
1175
+ self.device_mode = "AUTO"
1176
+ self.detection_hint = "center-1"
1177
+ self._sam_cache = {}
1178
+ self._bbox_cache = {}
1179
+
1180
+ @classmethod
1181
+ def INPUT_TYPES(s):
1182
+ bboxs = ["bbox/"+x for x in folder_paths.get_filename_list("ultralytics_bbox")]
1183
+ segms = ["segm/"+x for x in folder_paths.get_filename_list("ultralytics_segm")]
1184
+ sam_models = [x for x in folder_paths.get_filename_list("sams") if 'hq' not in x]
1185
+ return {
1186
+ "required": {
1187
+ "image": ("IMAGE",),
1188
+ "swapped_image": ("IMAGE",),
1189
+ "bbox_model_name": (bboxs + segms, ),
1190
+ "bbox_threshold": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}),
1191
+ "bbox_dilation": ("INT", {"default": 10, "min": -512, "max": 512, "step": 1}),
1192
+ "bbox_crop_factor": ("FLOAT", {"default": 3.0, "min": 1.0, "max": 100, "step": 0.1}),
1193
+ "bbox_drop_size": ("INT", {"min": 1, "max": 8192, "step": 1, "default": 10}),
1194
+ "sam_model_name": (sam_models, ),
1195
+ "sam_dilation": ("INT", {"default": 0, "min": -512, "max": 512, "step": 1}),
1196
+ "sam_threshold": ("FLOAT", {"default": 0.93, "min": 0.0, "max": 1.0, "step": 0.01}),
1197
+ "bbox_expansion": ("INT", {"default": 0, "min": 0, "max": 1000, "step": 1}),
1198
+ "mask_hint_threshold": ("FLOAT", {"default": 0.7, "min": 0.0, "max": 1.0, "step": 0.01}),
1199
+ "mask_hint_use_negative": (["False", "Small", "Outter"], ),
1200
+ "morphology_operation": (["dilate", "erode", "open", "close"],),
1201
+ "morphology_distance": ("INT", {"default": 0, "min": 0, "max": 128, "step": 1}),
1202
+ "blur_radius": ("INT", {"default": 9, "min": 0, "max": 48, "step": 1}),
1203
+ "sigma_factor": ("FLOAT", {"default": 1.0, "min": 0.01, "max": 3., "step": 0.01}),
1204
+ },
1205
+ "optional": {
1206
+ "mask_optional": ("MASK",),
1207
+ }
1208
+ }
1209
+
1210
+ RETURN_TYPES = ("IMAGE","MASK","IMAGE","IMAGE")
1211
+ RETURN_NAMES = ("IMAGE","MASK","MASK_PREVIEW","SWAPPED_FACE")
1212
+ FUNCTION = "execute"
1213
+ CATEGORY = "🌌 ReActor"
1214
+
1215
+ def execute(self, image, swapped_image, bbox_model_name, bbox_threshold, bbox_dilation, bbox_crop_factor, bbox_drop_size, sam_model_name, sam_dilation, sam_threshold, bbox_expansion, mask_hint_threshold, mask_hint_use_negative, morphology_operation, morphology_distance, blur_radius, sigma_factor, mask_optional=None):
1216
+ device = model_management.get_torch_device()
1217
+
1218
+ # ΠžΠΏΡ‚ΠΈΠΌΠ°Π»ΡŒΠ½ΠΎ ΠΏΠ΅Ρ€Π΅ΠΌΠ΅Ρ‰Π°Π΅ΠΌ Ρ‚Π΅Π½Π·ΠΎΡ€Ρ‹
1219
+ if isinstance(image, torch.Tensor) and image.device != device:
1220
+ image = image.to(device)
1221
+
1222
+ if isinstance(swapped_image, torch.Tensor) and swapped_image.device != device:
1223
+ swapped_image = swapped_image.to(device)
1224
+
1225
+ if mask_optional is not None:
1226
+ combined_mask = mask_optional
1227
+ else:
1228
+ # Load and cache BBox model
1229
+ if bbox_model_name not in self._bbox_cache:
1230
+ bbox_model_path = folder_paths.get_full_path("ultralytics", bbox_model_name)
1231
+ model = subcore.load_yolo(bbox_model_path)
1232
+ self._bbox_cache[bbox_model_name] = subcore.UltraBBoxDetector(model)
1233
+ bbox_detector = self._bbox_cache[bbox_model_name]
1234
+
1235
+ segs_all, seg_labels = bbox_detector.detect(image, bbox_threshold, bbox_dilation, bbox_crop_factor, bbox_drop_size, self.detailer_hook)
1236
+
1237
+ if self.labels != 'all':
1238
+ labels = self.labels.split(',') if isinstance(self.labels, str) else self.labels
1239
+ segs_all, _ = masking_segs.filter(segs_all, labels)
1240
+
1241
+ # Load and cache SAM model
1242
+ if sam_model_name not in self._sam_cache:
1243
+ sam_model_path = folder_paths.get_full_path("sams", sam_model_name)
1244
+ if 'vit_h' in sam_model_name:
1245
+ model_kind = 'vit_h'
1246
+ elif 'vit_l' in sam_model_name:
1247
+ model_kind = 'vit_l'
1248
+ else:
1249
+ model_kind = 'vit_b'
1250
+ sam = sam_model_registry[model_kind](checkpoint=sam_model_path)
1251
+ size = os.path.getsize(sam_model_path)
1252
+ sam.safe_to = core.SafeToGPU(size)
1253
+ sam.safe_to.to_device(sam, device)
1254
+ sam.is_auto_mode = self.device_mode == "AUTO"
1255
+ self._sam_cache[sam_model_name] = sam
1256
+ else:
1257
+ sam = self._sam_cache[sam_model_name]
1258
+
1259
+ # Handle batched input
1260
+ if image.ndim == 4:
1261
+ combined_masks = []
1262
+ for i in range(image.shape[0]):
1263
+ segs_i = segs_all[i] if i < len(segs_all) else []
1264
+ segs_tuple = ([segs_i], seg_labels) if isinstance(segs_i, dict) else (segs_i, seg_labels)
1265
+ image_device = image.to(device) if image.device != device else image
1266
+ image_i = image_device[i]
1267
+ mask_i, _ = core.make_sam_mask_segmented(
1268
+ sam, segs_tuple, image_i, self.detection_hint,
1269
+ sam_dilation, sam_threshold, bbox_expansion,
1270
+ mask_hint_threshold, mask_hint_use_negative
1271
+ )
1272
+ combined_masks.append(mask_i)
1273
+ combined_mask = torch.stack(combined_masks)
1274
+ else:
1275
+ image_device = image.to(device) if image.device != device else image
1276
+ combined_mask, _ = core.make_sam_mask_segmented(
1277
+ sam, (segs_all, seg_labels), image_device, self.detection_hint,
1278
+ sam_dilation, sam_threshold, bbox_expansion,
1279
+ mask_hint_threshold, mask_hint_use_negative
1280
+ )
1281
+
1282
+ # Morph operations
1283
+ if morphology_operation == "dilate":
1284
+ # print(f"max before: {combined_mask.max()}, min: {combined_mask.min()}, sum: {combined_mask.sum()}")
1285
+ combined_mask = self.iterative_morphology(combined_mask, morphology_distance, op="dilate")
1286
+ # print(f"after morph: {combined_mask.max()}, min: {combined_mask.min()}, sum: {combined_mask.sum()}")
1287
+ elif morphology_operation == "erode":
1288
+ combined_mask = self.iterative_morphology(combined_mask, morphology_distance, op="erode")
1289
+ elif morphology_operation == "open":
1290
+ combined_mask = self.iterative_morphology(self.iterative_morphology(combined_mask, morphology_distance, op="erode"), morphology_distance, op="dilate")
1291
+ elif morphology_operation == "close":
1292
+ combined_mask = self.iterative_morphology(self.iterative_morphology(combined_mask, morphology_distance, op="dilate"), morphology_distance, op="erode")
1293
+
1294
+ # Gaussian blur
1295
+ if blur_radius > 0:
1296
+ blur = T.GaussianBlur(kernel_size=blur_radius * 2 + 1, sigma=sigma_factor)
1297
+ mask_blurred = blur(combined_mask.unsqueeze(1)).squeeze(1)
1298
+ else:
1299
+ mask_blurred = combined_mask
1300
+
1301
+ # Apply mask to swapped image (basic RGBA composite)
1302
+ swapped_image = swapped_image.to(device) if swapped_image.device != device else swapped_image
1303
+ swapped_rgba = core.tensor2rgba(swapped_image)
1304
+
1305
+ mask_image_final = mask_blurred
1306
+
1307
+ # *** CUT BY MASK ***:
1308
+
1309
+ if len(swapped_image.shape) < 4:
1310
+ C = 1
1311
+ else:
1312
+ C = swapped_image.shape[3]
1313
+
1314
+ # We operate on RGBA to keep the code clean and then convert back after
1315
+ swapped_image = core.tensor2rgba(swapped_image)
1316
+ mask = core.tensor2mask(mask_image_final)
1317
+
1318
+ # Scale the mask to be a matching size if it isn't
1319
+ B, H, W, _ = swapped_image.shape
1320
+ mask = torch.nn.functional.interpolate(mask.unsqueeze(1), size=(H, W), mode='nearest')[:,0,:,:]
1321
+ MB, _, _ = mask.shape
1322
+
1323
+ if MB < B:
1324
+ assert(B % MB == 0)
1325
+ mask = mask.repeat(B // MB, 1, 1)
1326
+
1327
+ # masks_to_boxes errors if the tensor is all zeros, so we'll add a single pixel and zero it out at the end
1328
+ is_empty = ~torch.gt(torch.max(torch.reshape(mask,[MB, H * W]), dim=1).values, 0.)
1329
+ mask[is_empty,0,0] = 1.
1330
+ boxes = masks_to_boxes(mask)
1331
+ mask[is_empty,0,0] = 0.
1332
+
1333
+ min_x = boxes[:,0]
1334
+ min_y = boxes[:,1]
1335
+ max_x = boxes[:,2]
1336
+ max_y = boxes[:,3]
1337
+
1338
+ width = max_x - min_x + 1
1339
+ height = max_y - min_y + 1
1340
+
1341
+ use_width = int(torch.max(width).item())
1342
+ use_height = int(torch.max(height).item())
1343
+
1344
+ alpha_mask = torch.ones((B, H, W, 4))
1345
+ alpha_mask[:,:,:,3] = mask
1346
+
1347
+ alpha_mask = alpha_mask.to(device) if alpha_mask.device != device else alpha_mask
1348
+
1349
+ swapped_image = swapped_image * alpha_mask
1350
+
1351
+ cutted_image = torch.zeros((B, use_height, use_width, 4))
1352
+ for i in range(0, B):
1353
+ if not is_empty[i]:
1354
+ ymin = int(min_y[i].item())
1355
+ ymax = int(max_y[i].item())
1356
+ xmin = int(min_x[i].item())
1357
+ xmax = int(max_x[i].item())
1358
+ single = (swapped_image[i, ymin:ymax+1, xmin:xmax+1,:]).unsqueeze(0)
1359
+ resized = torch.nn.functional.interpolate(single.permute(0, 3, 1, 2), size=(use_height, use_width), mode='bicubic').permute(0, 2, 3, 1)
1360
+ cutted_image[i] = resized[0]
1361
+
1362
+ # Preserve our type unless we were previously RGB and added non-opaque alpha due to the mask size
1363
+ if C == 1:
1364
+ cutted_image = core.tensor2mask(cutted_image)
1365
+ elif C == 3 and torch.min(cutted_image[:,:,:,3]) == 1:
1366
+ cutted_image = core.tensor2rgb(cutted_image)
1367
+
1368
+ # *** PASTE BY MASK ***:
1369
+
1370
+ image_base = core.tensor2rgba(image)
1371
+ image_to_paste = core.tensor2rgba(cutted_image)
1372
+ mask = core.tensor2mask(mask_image_final)
1373
+
1374
+ # Scale the mask to be a matching size if it isn't
1375
+ B, H, W, C = image_base.shape
1376
+ MB = mask.shape[0]
1377
+ PB = image_to_paste.shape[0]
1378
+
1379
+ if B < PB:
1380
+ assert(PB % B == 0)
1381
+ image_base = image_base.repeat(PB // B, 1, 1, 1)
1382
+ B, H, W, C = image_base.shape
1383
+ if MB < B:
1384
+ assert(B % MB == 0)
1385
+ mask = mask.repeat(B // MB, 1, 1)
1386
+ elif B < MB:
1387
+ assert(MB % B == 0)
1388
+ image_base = image_base.repeat(MB // B, 1, 1, 1)
1389
+ if PB < B:
1390
+ assert(B % PB == 0)
1391
+ image_to_paste = image_to_paste.repeat(B // PB, 1, 1, 1)
1392
+
1393
+ mask = torch.nn.functional.interpolate(mask.unsqueeze(1), size=(H, W), mode='nearest')[:,0,:,:]
1394
+ MB, MH, MW = mask.shape
1395
+
1396
+ # masks_to_boxes errors if the tensor is all zeros, so we'll add a single pixel and zero it out at the end
1397
+ is_empty = ~torch.gt(torch.max(torch.reshape(mask,[MB, MH * MW]), dim=1).values, 0.)
1398
+ mask[is_empty,0,0] = 1.
1399
+ boxes = masks_to_boxes(mask)
1400
+ mask[is_empty,0,0] = 0.
1401
+
1402
+ min_x = boxes[:,0]
1403
+ min_y = boxes[:,1]
1404
+ max_x = boxes[:,2]
1405
+ max_y = boxes[:,3]
1406
+ mid_x = (min_x + max_x) / 2
1407
+ mid_y = (min_y + max_y) / 2
1408
+
1409
+ target_width = max_x - min_x + 1
1410
+ target_height = max_y - min_y + 1
1411
+
1412
+ result = image_base.detach().clone()
1413
+ face_segment = mask_image_final
1414
+
1415
+ pbar = progress_bar(MB)
1416
+
1417
+ for i in range(0, MB):
1418
+ if is_empty[i]:
1419
+ pbar.update(1)
1420
+ continue
1421
+ else:
1422
+ image_index = i
1423
+ SB, SH, SW, _ = image_to_paste.shape
1424
+
1425
+ # Figure out the desired size
1426
+ width = int(target_width[i].item())
1427
+ height = int(target_height[i].item())
1428
+
1429
+ width = SW
1430
+ height = SH
1431
+
1432
+ # Resize the image we're pasting if needed
1433
+ resized_image = image_to_paste[i].unsqueeze(0)
1434
+
1435
+ pasting = torch.ones([H, W, C])
1436
+ ymid = float(mid_y[i].item())
1437
+ ymin = int(math.floor(ymid - height / 2)) + 1
1438
+ ymax = int(math.floor(ymid + height / 2)) + 1
1439
+ xmid = float(mid_x[i].item())
1440
+ xmin = int(math.floor(xmid - width / 2)) + 1
1441
+ xmax = int(math.floor(xmid + width / 2)) + 1
1442
+
1443
+ _, source_ymax, source_xmax, _ = resized_image.shape
1444
+ source_ymin, source_xmin = 0, 0
1445
+
1446
+ if xmin < 0:
1447
+ source_xmin = abs(xmin)
1448
+ xmin = 0
1449
+ if ymin < 0:
1450
+ source_ymin = abs(ymin)
1451
+ ymin = 0
1452
+ if xmax > W:
1453
+ source_xmax -= (xmax - W)
1454
+ xmax = W
1455
+ if ymax > H:
1456
+ source_ymax -= (ymax - H)
1457
+ ymax = H
1458
+
1459
+ pasting[ymin:ymax, xmin:xmax, :] = resized_image[0, source_ymin:source_ymax, source_xmin:source_xmax, :]
1460
+ pasting[:, :, 3] = 1.
1461
+
1462
+ pasting_alpha = torch.zeros([H, W])
1463
+ pasting_alpha[ymin:ymax, xmin:xmax] = resized_image[0, source_ymin:source_ymax, source_xmin:source_xmax, 3]
1464
+
1465
+ paste_mask = torch.min(pasting_alpha, mask[i]).unsqueeze(2).repeat(1, 1, 4)
1466
+
1467
+ pasting = pasting.to(device) if pasting.device != device else pasting
1468
+ paste_mask = paste_mask.to(device) if paste_mask.device != device else paste_mask
1469
+
1470
+ result[image_index] = pasting * paste_mask + result[image_index] * (1. - paste_mask)
1471
+
1472
+ face_segment = result
1473
+
1474
+ face_segment[...,3] = mask[i]
1475
+
1476
+ result = rgba2rgb_tensor(result)
1477
+ result = result.cpu() # ΠŸΠ΅Ρ€Π΅ΠΌΠ΅Ρ‰Π°Π΅ΠΌ Ρ€Π΅Π·ΡƒΠ»ΡŒΡ‚Π°Ρ‚ ΠΎΠ±Ρ€Π°Ρ‚Π½ΠΎ Π½Π° CPU
1478
+
1479
+ pbar.update(1)
1480
+
1481
+ try:
1482
+ torch.cuda.empty_cache()
1483
+ except:
1484
+ pass
1485
+
1486
+ progress_bar_reset(pbar)
1487
+
1488
+ return (result, combined_mask, mask_blurred, face_segment)
1489
+
1490
+ def iterative_morphology(self, image, distance, op="dilate"):
1491
+ if distance <= 0:
1492
+ return image
1493
+ image = image.unsqueeze(1) # shape [B, 1, H, W] or [1, 1, H, W]
1494
+ for _ in range(distance):
1495
+ if op == "dilate":
1496
+ image = F.max_pool2d(image, kernel_size=3, stride=1, padding=1)
1497
+ elif op == "erode":
1498
+ image = -F.max_pool2d(-image, kernel_size=3, stride=1, padding=1)
1499
+ return image.squeeze(1)
1500
+
1501
+
1502
+ class ImageDublicator:
1503
+ @classmethod
1504
+ def INPUT_TYPES(s):
1505
+ return {
1506
+ "required": {
1507
+ "image": ("IMAGE",),
1508
+ "count": ("INT", {"default": 1, "min": 0}),
1509
+ },
1510
+ }
1511
+
1512
+ RETURN_TYPES = ("IMAGE",)
1513
+ RETURN_NAMES = ("IMAGES",)
1514
+ OUTPUT_IS_LIST = (True,)
1515
+ FUNCTION = "execute"
1516
+ CATEGORY = "🌌 ReActor"
1517
+
1518
+ def execute(self, image, count):
1519
+ images = [image for i in range(count)]
1520
+ return (images,)
1521
+
1522
+
1523
+ class ImageRGBA2RGB:
1524
+ @classmethod
1525
+ def INPUT_TYPES(s):
1526
+ return {
1527
+ "required": {
1528
+ "image": ("IMAGE",),
1529
+ },
1530
+ }
1531
+
1532
+ RETURN_TYPES = ("IMAGE",)
1533
+ FUNCTION = "execute"
1534
+ CATEGORY = "🌌 ReActor"
1535
+
1536
+ def execute(self, image):
1537
+ out = rgba2rgb_tensor(image)
1538
+ return (out,)
1539
+
1540
+
1541
+ class MakeFaceModelBatch:
1542
+ @classmethod
1543
+ def INPUT_TYPES(s):
1544
+ return {
1545
+ "required": {
1546
+ "face_model1": ("FACE_MODEL",),
1547
+ },
1548
+ "optional": {
1549
+ "face_model2": ("FACE_MODEL",),
1550
+ "face_model3": ("FACE_MODEL",),
1551
+ "face_model4": ("FACE_MODEL",),
1552
+ "face_model5": ("FACE_MODEL",),
1553
+ "face_model6": ("FACE_MODEL",),
1554
+ "face_model7": ("FACE_MODEL",),
1555
+ "face_model8": ("FACE_MODEL",),
1556
+ "face_model9": ("FACE_MODEL",),
1557
+ "face_model10": ("FACE_MODEL",),
1558
+ },
1559
+ }
1560
+
1561
+ RETURN_TYPES = ("FACE_MODEL",)
1562
+ RETURN_NAMES = ("FACE_MODELS",)
1563
+ FUNCTION = "execute"
1564
+
1565
+ CATEGORY = "🌌 ReActor"
1566
+
1567
+ def execute(self, **kwargs):
1568
+ if len(kwargs) > 0:
1569
+ face_models = [value for value in kwargs.values()]
1570
+ return (face_models,)
1571
+ else:
1572
+ logger.error("Please provide at least 1 `face_model`")
1573
+ return (None,)
1574
+
1575
+
1576
+ class ReActorOptions:
1577
+ @classmethod
1578
+ def INPUT_TYPES(s):
1579
+ return {
1580
+ "required": {
1581
+ "input_faces_order": (
1582
+ ["left-right","right-left","top-bottom","bottom-top","small-large","large-small"], {"default": "large-small"}
1583
+ ),
1584
+ "input_faces_index": ("STRING", {"default": "0"}),
1585
+ "detect_gender_input": (["no","female","male"], {"default": "no"}),
1586
+ "source_faces_order": (
1587
+ ["left-right","right-left","top-bottom","bottom-top","small-large","large-small"], {"default": "large-small"}
1588
+ ),
1589
+ "source_faces_index": ("STRING", {"default": "0"}),
1590
+ "detect_gender_source": (["no","female","male"], {"default": "no"}),
1591
+ "console_log_level": ([0, 1, 2], {"default": 1}),
1592
+ "restore_swapped_only": ("BOOLEAN", {"default": True, "label_off": "no", "label_on": "yes"})
1593
+ }
1594
+ }
1595
+
1596
+ RETURN_TYPES = ("OPTIONS",)
1597
+ FUNCTION = "execute"
1598
+ CATEGORY = "🌌 ReActor"
1599
+
1600
+ def execute(self,input_faces_order, input_faces_index, detect_gender_input, source_faces_order, source_faces_index, detect_gender_source, console_log_level, restore_swapped_only):
1601
+ options: dict = {
1602
+ "input_faces_order": input_faces_order,
1603
+ "input_faces_index": input_faces_index,
1604
+ "detect_gender_input": detect_gender_input,
1605
+ "source_faces_order": source_faces_order,
1606
+ "source_faces_index": source_faces_index,
1607
+ "detect_gender_source": detect_gender_source,
1608
+ "console_log_level": console_log_level,
1609
+ "restore_swapped_only": restore_swapped_only,
1610
+ }
1611
+ return (options, )
1612
+
1613
+
1614
+ class ReActorFaceBoost:
1615
+ @classmethod
1616
+ def INPUT_TYPES(s):
1617
+ return {
1618
+ "required": {
1619
+ "enabled": ("BOOLEAN", {"default": True, "label_off": "OFF", "label_on": "ON"}),
1620
+ "boost_model": (get_model_names(get_restorers),),
1621
+ "interpolation": (["Nearest","Bilinear","Bicubic","Lanczos"], {"default": "Bicubic"}),
1622
+ "visibility": ("FLOAT", {"default": 1, "min": 0.1, "max": 1, "step": 0.05}),
1623
+ "codeformer_weight": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1, "step": 0.05}),
1624
+ "restore_with_main_after": ("BOOLEAN", {"default": False}),
1625
+ }
1626
+ }
1627
+
1628
+ RETURN_TYPES = ("FACE_BOOST",)
1629
+ FUNCTION = "execute"
1630
+ CATEGORY = "🌌 ReActor"
1631
+
1632
+ def execute(self,enabled,boost_model,interpolation,visibility,codeformer_weight,restore_with_main_after):
1633
+ face_boost: dict = {
1634
+ "enabled": enabled,
1635
+ "boost_model": boost_model,
1636
+ "interpolation": interpolation,
1637
+ "visibility": visibility,
1638
+ "codeformer_weight": codeformer_weight,
1639
+ "restore_with_main_after": restore_with_main_after,
1640
+ }
1641
+ return (face_boost, )
1642
+
1643
+ class ReActorUnload:
1644
+ @classmethod
1645
+ def INPUT_TYPES(s):
1646
+ return {
1647
+ "required": {
1648
+ "trigger": ("IMAGE", ),
1649
+ },
1650
+ }
1651
+
1652
+ RETURN_TYPES = ("IMAGE",)
1653
+ FUNCTION = "execute"
1654
+ CATEGORY = "🌌 ReActor"
1655
+
1656
+ def execute(self, trigger):
1657
+ unload_all_models()
1658
+ return (trigger,)
1659
+
1660
+
1661
+ NODE_CLASS_MAPPINGS = {
1662
+ # --- MAIN NODES ---
1663
+ "ReActorFaceSwap": reactor,
1664
+ "ReActorFaceSwapOpt": ReActorPlusOpt,
1665
+ "ReActorOptions": ReActorOptions,
1666
+ "ReActorFaceBoost": ReActorFaceBoost,
1667
+ "ReActorMaskHelper": MaskHelper,
1668
+ "ReActorSetWeight": ReActorWeight,
1669
+ # --- Operations with Face Models ---
1670
+ "ReActorSaveFaceModel": SaveFaceModel,
1671
+ "ReActorLoadFaceModel": LoadFaceModel,
1672
+ "ReActorBuildFaceModel": BuildFaceModel,
1673
+ "ReActorMakeFaceModelBatch": MakeFaceModelBatch,
1674
+ # --- Additional Nodes ---
1675
+ "ReActorRestoreFace": RestoreFace,
1676
+ "ReActorRestoreFaceAdvanced": RestoreFaceAdvanced,
1677
+ "ReActorImageDublicator": ImageDublicator,
1678
+ "ImageRGBA2RGB": ImageRGBA2RGB,
1679
+ "ReActorUnload": ReActorUnload,
1680
+ }
1681
+
1682
+ NODE_DISPLAY_NAME_MAPPINGS = {
1683
+ # --- MAIN NODES ---
1684
+ "ReActorFaceSwap": "ReActor 🌌 Fast Face Swap",
1685
+ "ReActorFaceSwapOpt": "ReActor 🌌 Fast Face Swap [OPTIONS]",
1686
+ "ReActorOptions": "ReActor 🌌 Options",
1687
+ "ReActorFaceBoost": "ReActor 🌌 Face Booster",
1688
+ "ReActorMaskHelper": "ReActor 🌌 Masking Helper",
1689
+ "ReActorSetWeight": "ReActor 🌌 Set Face Swap Weight",
1690
+ # --- Operations with Face Models ---
1691
+ "ReActorSaveFaceModel": "Save Face Model 🌌 ReActor",
1692
+ "ReActorLoadFaceModel": "Load Face Model 🌌 ReActor",
1693
+ "ReActorBuildFaceModel": "Build Blended Face Model 🌌 ReActor",
1694
+ "ReActorMakeFaceModelBatch": "Make Face Model Batch 🌌 ReActor",
1695
+ # --- Additional Nodes ---
1696
+ "ReActorRestoreFace": "Restore Face 🌌 ReActor",
1697
+ "ReActorRestoreFaceAdvanced": "Restore Face Advanced 🌌 ReActor",
1698
+ "ReActorImageDublicator": "Image Dublicator (List) 🌌 ReActor",
1699
+ "ImageRGBA2RGB": "Convert RGBA to RGB 🌌 ReActor",
1700
+ "ReActorUnload": "Unload ReActor Models 🌌 ReActor",
1701
+ }