File size: 12,047 Bytes
983684c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1c60403
d176e27
983684c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a6c71e9
983684c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e9ea181
983684c
 
 
 
 
13706ea
998a434
13706ea
998a434
dbf6cac
983684c
 
 
 
 
 
998a434
983684c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
from __future__ import annotations
import gradio as gr
import pathlib
import sys
sys.path.insert(0, 'vtoonify')

from util import load_psp_standalone, get_video_crop_parameter, tensor2cv2
import torch
import torch.nn as nn
import numpy as np
import dlib
import cv2
from model.vtoonify import VToonify
from model.bisenet.model import BiSeNet
import torch.nn.functional as F
from torchvision import transforms
from model.encoder.align_all_parallel import align_face
import gc
import huggingface_hub
import os

MODEL_REPO = 'PKUWilliamYang/VToonify'

class Model():
    def __init__(self, device):
        super().__init__()
        
        self.device = device
        self.style_types = {
            'cartoon1': ['vtoonify_d_cartoon/vtoonify_s026_d0.5.pt', 26],
            'cartoon1-d': ['vtoonify_d_cartoon/vtoonify_s_d.pt', 26],
            'cartoon2-d': ['vtoonify_d_cartoon/vtoonify_s_d.pt', 64],
            'cartoon3-d': ['vtoonify_d_cartoon/vtoonify_s_d.pt', 153],
            'cartoon4': ['vtoonify_d_cartoon/vtoonify_s299_d0.5.pt', 299],
            'cartoon4-d': ['vtoonify_d_cartoon/vtoonify_s_d.pt', 299],
            'cartoon5-d': ['vtoonify_d_cartoon/vtoonify_s_d.pt', 8],
            'comic1-d': ['vtoonify_d_comic/vtoonify_s_d.pt', 28],
            'comic2-d': ['vtoonify_d_comic/vtoonify_s_d.pt', 18],
            'comic3-d': ['vtoonify_d_illustration/vtoonify_s054_d_c.pt', 54],
            'arcane1': ['vtoonify_d_arcane/vtoonify_s000_d0.5.pt', 0],
            'arcane1-d': ['vtoonify_d_arcane/vtoonify_s_d.pt', 0],
            'arcane2': ['vtoonify_d_arcane/vtoonify_s077_d0.5.pt', 77],
            'arcane2-d': ['vtoonify_d_arcane/vtoonify_s_d.pt', 77],
            'caricature1': ['vtoonify_d_caricature/vtoonify_s039_d0.5.pt', 39],
            'caricature2': ['vtoonify_d_caricature/vtoonify_s068_d0.5.pt', 68],
            'pixar': ['vtoonify_d_pixar/vtoonify_s052_d0.5.pt', 52],
            'pixar-d': ['vtoonify_d_pixar/vtoonify_s_d.pt', 52],
        }
        
        self.landmarkpredictor = self._create_dlib_landmark_model()
        self.parsingpredictor = self._create_parsing_model()
        self.pspencoder = self._load_encoder()    
        self.transform = transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.5, 0.5, 0.5],std=[0.5,0.5,0.5]),
            ])
        
        self.vtoonify, self.exstyle = self._load_default_model()
        self.color_transfer = False
        
    @staticmethod
    def _create_dlib_landmark_model():
        return dlib.shape_predictor(huggingface_hub.hf_hub_download(MODEL_REPO,
                                                    'models/shape_predictor_68_face_landmarks.dat'))
    
    def _create_parsing_model(self):
        parsingpredictor = BiSeNet(n_classes=19)
        parsingpredictor.load_state_dict(torch.load(huggingface_hub.hf_hub_download(MODEL_REPO, 'models/faceparsing.pth'),
                                                    map_location=lambda storage, loc: storage))
        parsingpredictor.to(self.device).eval()
        return parsingpredictor
    
    def _load_encoder(self) -> nn.Module:
        style_encoder_path = huggingface_hub.hf_hub_download(MODEL_REPO,'models/encoder.pt')
        return load_psp_standalone(style_encoder_path, self.device)
    
    def _load_default_model(self) -> tuple[torch.Tensor, str]:
        vtoonify = VToonify(backbone = 'dualstylegan')
        vtoonify.load_state_dict(torch.load(huggingface_hub.hf_hub_download(MODEL_REPO,
                                            'models/vtoonify_d_cartoon/vtoonify_s026_d0.5.pt'), 
                                            map_location=lambda storage, loc: storage)['g_ema'])
        vtoonify.to(self.device)
        tmp = np.load(huggingface_hub.hf_hub_download(MODEL_REPO,'models/vtoonify_d_cartoon/exstyle_code.npy'), allow_pickle=True).item()
        exstyle = torch.tensor(tmp[list(tmp.keys())[26]]).to(self.device)
        with torch.no_grad():  
            exstyle = vtoonify.zplus2wplus(exstyle)
        return vtoonify, exstyle
    
    def load_model(self, style_type: str) -> tuple[torch.Tensor, str]:
        if style_type == 'comic3-d':
            self.color_transfer = True
        else:
            self.color_transfer = False
        model_path, ind = self.style_types[style_type]
        style_path = os.path.join('models',os.path.dirname(model_path),'exstyle_code.npy')
        self.vtoonify.load_state_dict(torch.load(huggingface_hub.hf_hub_download(MODEL_REPO,'models/'+model_path), 
                                            map_location=lambda storage, loc: storage)['g_ema'])
        tmp = np.load(huggingface_hub.hf_hub_download(MODEL_REPO, style_path), allow_pickle=True).item()
        exstyle = torch.tensor(tmp[list(tmp.keys())[ind]]).to(self.device)
        with torch.no_grad():  
            exstyle = self.vtoonify.zplus2wplus(exstyle)
        return exstyle, 'Model of %s loaded.'%(style_type)
    
    def detect_and_align(self, frame, top, bottom, left, right, return_para=False):
        message = 'Error: no face detected!'
        paras = get_video_crop_parameter(frame, self.landmarkpredictor, [left, right, top, bottom])
        instyle = torch.zeros(1,18,512).to(self.device)
        h, w, scale = 0, 0, 0
        if paras is not None:
            h,w,top,bottom,left,right,scale = paras
            H, W = int(bottom-top), int(right-left)
            # for HR image, we apply gaussian blur to it to avoid over-sharp stylization results
            kernel_1d = np.array([[0.125],[0.375],[0.375],[0.125]])
            if scale <= 0.75:
                frame = cv2.sepFilter2D(frame, -1, kernel_1d, kernel_1d)
            if scale <= 0.375:
                frame = cv2.sepFilter2D(frame, -1, kernel_1d, kernel_1d)
            frame = cv2.resize(frame, (w, h))[top:bottom, left:right]
            with torch.no_grad():
                I = align_face(frame, self.landmarkpredictor)
                if I is not None:
                    I = self.transform(I).unsqueeze(dim=0).to(self.device)
                    instyle = self.pspencoder(I)
                    instyle = self.vtoonify.zplus2wplus(instyle)
                    message = 'Successfully rescale the frame to (%d, %d)'%(bottom-top, right-left)
                else:
                    frame = np.zeros((256,256,3), np.uint8)
        else:
            frame = np.zeros((256,256,3), np.uint8)
        if return_para:
            return frame, instyle, message, w, h, top, bottom, left, right, scale
        return frame, instyle, message
    
    #@torch.inference_mode()
    def detect_and_align_image(self, image: str, top: int, bottom: int, left: int, right: int
                              ) -> tuple[np.ndarray, torch.Tensor, str]:
        
        frame = cv2.imread(image)
        frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
        return self.detect_and_align(frame, top, bottom, left, right)
    
    def detect_and_align_video(self, video: str, top: int, bottom: int, left: int, right: int
                              ) -> tuple[np.ndarray, torch.Tensor, str]:

        video_cap = cv2.VideoCapture(video)
        success, frame = video_cap.read()
        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        video_cap.release()
        return self.detect_and_align(frame, top, bottom, left, right)
    
    def detect_and_align_full_video(self, video: str, top: int, bottom: int, left: int, right: int) -> tuple[str, torch.Tensor, str]:
        message = 'Error: no face detected!'
        instyle = torch.zeros(1,18,512).to(self.device)
        video_cap = cv2.VideoCapture(video)
        num = int(video_cap.get(7))
        success, frame = video_cap.read()
        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        frame, instyle, message, w, h, top, bottom, left, right, scale = self.detect_and_align(frame, top, bottom, left, right, True)
        if instyle is None:
            return 'default.mp4', instyle, message    
        fourcc = cv2.VideoWriter_fourcc(*'mp4v')
        videoWriter = cv2.VideoWriter('input.mp4', fourcc, video_cap.get(5), (int(right-left), int(bottom-top)))
        kernel_1d = np.array([[0.125],[0.375],[0.375],[0.125]])
        for i in range(num-1):
            success, frame = video_cap.read()
            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)            
            if scale <= 0.75:
                frame = cv2.sepFilter2D(frame, -1, kernel_1d, kernel_1d)
            if scale <= 0.375:
                frame = cv2.sepFilter2D(frame, -1, kernel_1d, kernel_1d)
            frame = cv2.resize(frame, (w, h))[top:bottom, left:right]
            videoWriter.write(cv2.cvtColor(frame, cv2.COLOR_RGB2BGR))

        videoWriter.release()
        video_cap.release()

        return 'input.mp4', instyle, 'Successfully rescale the video to (%d, %d)'%(bottom-top, right-left)
    
    def image_toonify(self, aligned_face: np.ndarray, instyle: torch.Tensor, exstyle: torch.Tensor, style_degree: float) -> np.ndarray:  
        if exstyle is None:
            exstyle = self.exstyle
        with torch.no_grad():
            if self.color_transfer:
                s_w = exstyle
            else:
                s_w = instyle.clone()
                s_w[:,:7] = exstyle[:,:7]

            x = self.transform(aligned_face).unsqueeze(dim=0).to(self.device)
            x_p = F.interpolate(self.parsingpredictor(2*(F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)))[0], 
                                scale_factor=0.5, recompute_scale_factor=False).detach()
            inputs = torch.cat((x, x_p/16.), dim=1)
            y_tilde = self.vtoonify(inputs, s_w.repeat(inputs.size(0), 1, 1), d_s = style_degree)        
            y_tilde = torch.clamp(y_tilde, -1, 1)

        return ((y_tilde[0].cpu().numpy().transpose(1, 2, 0) + 1.0) * 127.5).astype(np.uint8)
    
    def video_tooniy(self, aligned_video: str, instyle: torch.Tensor, exstyle: torch.Tensor, style_degree: float) -> str: 
        if exstyle is None:
            exstyle = self.exstyle
        video_cap = cv2.VideoCapture(aligned_video)
        num = int(video_cap.get(7))
        fourcc = cv2.VideoWriter_fourcc(*'mp4v')
        videoWriter = cv2.VideoWriter('output.mp4', fourcc, 
                                      video_cap.get(5), (int(video_cap.get(3)*4),
                                      int(video_cap.get(4)*4)))        

        batch_frames = []
        if video_cap.get(3) != 0:
            batch_size = max(1, int(round(4 * 256* 256/ video_cap.get(3) / video_cap.get(4))))
        else:
            batch_size = 1
        print('Using batch size of %d on %d frames'%(batch_size, num))
        with torch.no_grad():
            if self.color_transfer:
                s_w = exstyle
            else:
                s_w = instyle.clone()
                s_w[:,:7] = exstyle[:,:7]
            for i in range(num):
                success, frame = video_cap.read()
                frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                batch_frames += [self.transform(frame).unsqueeze(dim=0).to(self.device)]
                if len(batch_frames) == batch_size or (i+1) == num:
                    x = torch.cat(batch_frames, dim=0)
                    batch_frames = []
                    with torch.no_grad():
                        x_p = F.interpolate(self.parsingpredictor(2*(F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)))[0], 
                                            scale_factor=0.5, recompute_scale_factor=False).detach()
                        inputs = torch.cat((x, x_p/16.), dim=1)
                        y_tilde = self.vtoonify(inputs, s_w.repeat(inputs.size(0), 1, 1), style_degree)       
                        y_tilde = torch.clamp(y_tilde, -1, 1)
                    for k in range(y_tilde.size(0)):
                        videoWriter.write(tensor2cv2(y_tilde[k].cpu()))
                    gc.collect()

        videoWriter.release()
        video_cap.release()
        return 'output.mp4'