Spaces:
Running
on
Zero
Running
on
Zero
SunderAli17
commited on
Commit
•
f458bd8
1
Parent(s):
be0d2f2
Upload 3 files
Browse files- toonmage/fluxpipeline.py +188 -0
- toonmage/pipeline.py +232 -0
- toonmage/utils.py +76 -0
toonmage/fluxpipeline.py
ADDED
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gc
|
2 |
+
|
3 |
+
import cv2
|
4 |
+
import insightface
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
from basicsr.utils import img2tensor, tensor2img
|
8 |
+
from facexlib.parsing import init_parsing_model
|
9 |
+
from facexlib.utils.face_restoration_helper import FaceRestoreHelper
|
10 |
+
from huggingface_hub import hf_hub_download, snapshot_download
|
11 |
+
from insightface.app import FaceAnalysis
|
12 |
+
from safetensors.torch import load_file
|
13 |
+
from torchvision.transforms import InterpolationMode
|
14 |
+
from torchvision.transforms.functional import normalize, resize
|
15 |
+
|
16 |
+
from eva_clip import create_model_and_transforms
|
17 |
+
from eva_clip.constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
|
18 |
+
from toonmage.encoders_flux import IDFormer, PerceiverAttentionCA
|
19 |
+
|
20 |
+
|
21 |
+
class ToonMagePipeline(nn.Module):
|
22 |
+
def __init__(self, dit, device, weight_dtype=torch.bfloat16, *args, **kwargs):
|
23 |
+
super().__init__()
|
24 |
+
self.device = device
|
25 |
+
self.weight_dtype = weight_dtype
|
26 |
+
double_interval = 2
|
27 |
+
single_interval = 4
|
28 |
+
|
29 |
+
# init encoder
|
30 |
+
self.toonmage_encoder = IDFormer().to(self.device, self.weight_dtype)
|
31 |
+
|
32 |
+
num_ca = 19 // double_interval + 38 // single_interval
|
33 |
+
if 19 % double_interval != 0:
|
34 |
+
num_ca += 1
|
35 |
+
if 38 % single_interval != 0:
|
36 |
+
num_ca += 1
|
37 |
+
self.toonmage_ca = nn.ModuleList([
|
38 |
+
PerceiverAttentionCA().to(self.device, self.weight_dtype) for _ in range(num_ca)
|
39 |
+
])
|
40 |
+
|
41 |
+
dit.toonmage_ca = self.toonmage_ca
|
42 |
+
dit.toonmage_double_interval = double_interval
|
43 |
+
dit.toonmage_single_interval = single_interval
|
44 |
+
|
45 |
+
# preprocessors
|
46 |
+
# face align and parsing
|
47 |
+
self.face_helper = FaceRestoreHelper(
|
48 |
+
upscale_factor=1,
|
49 |
+
face_size=512,
|
50 |
+
crop_ratio=(1, 1),
|
51 |
+
det_model='retinaface_resnet50',
|
52 |
+
save_ext='png',
|
53 |
+
device=self.device,
|
54 |
+
)
|
55 |
+
self.face_helper.face_parse = None
|
56 |
+
self.face_helper.face_parse = init_parsing_model(model_name='bisenet', device=self.device)
|
57 |
+
# clip-vit backbone
|
58 |
+
model, _, _ = create_model_and_transforms('EVA02-CLIP-L-14-336', 'eva_clip', force_custom_clip=True)
|
59 |
+
model = model.visual
|
60 |
+
self.clip_vision_model = model.to(self.device, dtype=self.weight_dtype)
|
61 |
+
eva_transform_mean = getattr(self.clip_vision_model, 'image_mean', OPENAI_DATASET_MEAN)
|
62 |
+
eva_transform_std = getattr(self.clip_vision_model, 'image_std', OPENAI_DATASET_STD)
|
63 |
+
if not isinstance(eva_transform_mean, (list, tuple)):
|
64 |
+
eva_transform_mean = (eva_transform_mean,) * 3
|
65 |
+
if not isinstance(eva_transform_std, (list, tuple)):
|
66 |
+
eva_transform_std = (eva_transform_std,) * 3
|
67 |
+
self.eva_transform_mean = eva_transform_mean
|
68 |
+
self.eva_transform_std = eva_transform_std
|
69 |
+
# antelopev2
|
70 |
+
snapshot_download('DIAMONIK7777/antelopev2', local_dir='models/antelopev2')
|
71 |
+
self.app = FaceAnalysis(
|
72 |
+
name='antelopev2', root='.', providers=['CPUExecutionProvider']
|
73 |
+
)
|
74 |
+
self.app.prepare(ctx_id=0, det_size=(640, 640))
|
75 |
+
self.handler_ante = insightface.model_zoo.get_model('models/antelopev2/glintr100.onnx', providers=['CPUExecutionProvider'])
|
76 |
+
self.handler_ante.prepare(ctx_id=0)
|
77 |
+
|
78 |
+
gc.collect()
|
79 |
+
torch.cuda.empty_cache()
|
80 |
+
|
81 |
+
# self.load_pretrain()
|
82 |
+
|
83 |
+
# other configs
|
84 |
+
self.debug_img_list = []
|
85 |
+
|
86 |
+
def load_pretrain(self, pretrain_path=None):
|
87 |
+
hf_hub_download('SunderAli17/SAK', 'toonmage_flux_v2.safetensors', local_dir='models')
|
88 |
+
ckpt_path = 'models/toonmage_flux_v2.safetensors'
|
89 |
+
if pretrain_path is not None:
|
90 |
+
ckpt_path = pretrain_path
|
91 |
+
state_dict = load_file(ckpt_path)
|
92 |
+
state_dict_dict = {}
|
93 |
+
for k, v in state_dict.items():
|
94 |
+
module = k.split('.')[0]
|
95 |
+
state_dict_dict.setdefault(module, {})
|
96 |
+
new_k = k[len(module) + 1:]
|
97 |
+
state_dict_dict[module][new_k] = v
|
98 |
+
|
99 |
+
for module in state_dict_dict:
|
100 |
+
print(f'loading from {module}')
|
101 |
+
getattr(self, module).load_state_dict(state_dict_dict[module], strict=True)
|
102 |
+
|
103 |
+
del state_dict
|
104 |
+
del state_dict_dict
|
105 |
+
|
106 |
+
def to_gray(self, img):
|
107 |
+
x = 0.299 * img[:, 0:1] + 0.587 * img[:, 1:2] + 0.114 * img[:, 2:3]
|
108 |
+
x = x.repeat(1, 3, 1, 1)
|
109 |
+
return x
|
110 |
+
|
111 |
+
def get_id_embedding(self, image, cal_uncond=False):
|
112 |
+
"""
|
113 |
+
Args:
|
114 |
+
image: numpy rgb image, range [0, 255]
|
115 |
+
"""
|
116 |
+
self.face_helper.clean_all()
|
117 |
+
self.debug_img_list = []
|
118 |
+
image_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
|
119 |
+
# get antelopev2 embedding
|
120 |
+
# for k in self.app.models.keys():
|
121 |
+
# self.app.models[k].session.set_providers(['CUDAExecutionProvider'])
|
122 |
+
face_info = self.app.get(image_bgr)
|
123 |
+
if len(face_info) > 0:
|
124 |
+
face_info = sorted(face_info, key=lambda x: (x['bbox'][2] - x['bbox'][0]) * (x['bbox'][3] - x['bbox'][1]))[
|
125 |
+
-1
|
126 |
+
] # only use the maximum face
|
127 |
+
id_ante_embedding = face_info['embedding']
|
128 |
+
self.debug_img_list.append(
|
129 |
+
image[
|
130 |
+
int(face_info['bbox'][1]) : int(face_info['bbox'][3]),
|
131 |
+
int(face_info['bbox'][0]) : int(face_info['bbox'][2]),
|
132 |
+
]
|
133 |
+
)
|
134 |
+
else:
|
135 |
+
id_ante_embedding = None
|
136 |
+
|
137 |
+
# using facexlib to detect and align face
|
138 |
+
self.face_helper.read_image(image_bgr)
|
139 |
+
self.face_helper.get_face_landmarks_5(only_center_face=True)
|
140 |
+
self.face_helper.align_warp_face()
|
141 |
+
if len(self.face_helper.cropped_faces) == 0:
|
142 |
+
raise RuntimeError('facexlib align face fail')
|
143 |
+
align_face = self.face_helper.cropped_faces[0]
|
144 |
+
# incase insightface didn't detect face
|
145 |
+
if id_ante_embedding is None:
|
146 |
+
print('fail to detect face using insightface, extract embedding on align face')
|
147 |
+
# self.handler_ante.session.set_providers(['CUDAExecutionProvider'])
|
148 |
+
id_ante_embedding = self.handler_ante.get_feat(align_face)
|
149 |
+
|
150 |
+
id_ante_embedding = torch.from_numpy(id_ante_embedding).to(self.device, self.weight_dtype)
|
151 |
+
if id_ante_embedding.ndim == 1:
|
152 |
+
id_ante_embedding = id_ante_embedding.unsqueeze(0)
|
153 |
+
|
154 |
+
# parsing
|
155 |
+
input = img2tensor(align_face, bgr2rgb=True).unsqueeze(0) / 255.0
|
156 |
+
input = input.to(self.device)
|
157 |
+
parsing_out = self.face_helper.face_parse(normalize(input, [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]))[0]
|
158 |
+
parsing_out = parsing_out.argmax(dim=1, keepdim=True)
|
159 |
+
bg_label = [0, 16, 18, 7, 8, 9, 14, 15]
|
160 |
+
bg = sum(parsing_out == i for i in bg_label).bool()
|
161 |
+
white_image = torch.ones_like(input)
|
162 |
+
# only keep the face features
|
163 |
+
face_features_image = torch.where(bg, white_image, self.to_gray(input))
|
164 |
+
self.debug_img_list.append(tensor2img(face_features_image, rgb2bgr=False))
|
165 |
+
|
166 |
+
# transform img before sending to eva-clip-vit
|
167 |
+
face_features_image = resize(face_features_image, self.clip_vision_model.image_size, InterpolationMode.BICUBIC)
|
168 |
+
face_features_image = normalize(face_features_image, self.eva_transform_mean, self.eva_transform_std)
|
169 |
+
id_cond_vit, id_vit_hidden = self.clip_vision_model(
|
170 |
+
face_features_image.to(self.weight_dtype), return_all_features=False, return_hidden=True, shuffle=False
|
171 |
+
)
|
172 |
+
id_cond_vit_norm = torch.norm(id_cond_vit, 2, 1, True)
|
173 |
+
id_cond_vit = torch.div(id_cond_vit, id_cond_vit_norm)
|
174 |
+
|
175 |
+
id_cond = torch.cat([id_ante_embedding, id_cond_vit], dim=-1)
|
176 |
+
|
177 |
+
id_embedding = self.toonmage_encoder(id_cond, id_vit_hidden)
|
178 |
+
|
179 |
+
if not cal_uncond:
|
180 |
+
return id_embedding, None
|
181 |
+
|
182 |
+
id_uncond = torch.zeros_like(id_cond)
|
183 |
+
id_vit_hidden_uncond = []
|
184 |
+
for layer_idx in range(0, len(id_vit_hidden)):
|
185 |
+
id_vit_hidden_uncond.append(torch.zeros_like(id_vit_hidden[layer_idx]))
|
186 |
+
uncond_id_embedding = self.toonmage_encoder(id_uncond, id_vit_hidden_uncond)
|
187 |
+
|
188 |
+
return id_embedding, uncond_id_embedding
|
toonmage/pipeline.py
ADDED
@@ -0,0 +1,232 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gc
|
2 |
+
|
3 |
+
import cv2
|
4 |
+
import insightface
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
from basicsr.utils import img2tensor, tensor2img
|
8 |
+
from diffusers import (
|
9 |
+
DPMSolverMultistepScheduler,
|
10 |
+
StableDiffusionXLPipeline,
|
11 |
+
UNet2DConditionModel,
|
12 |
+
)
|
13 |
+
from facexlib.parsing import init_parsing_model
|
14 |
+
from facexlib.utils.face_restoration_helper import FaceRestoreHelper
|
15 |
+
from huggingface_hub import hf_hub_download, snapshot_download
|
16 |
+
from insightface.app import FaceAnalysis
|
17 |
+
from safetensors.torch import load_file
|
18 |
+
from torchvision.transforms import InterpolationMode
|
19 |
+
from torchvision.transforms.functional import normalize, resize
|
20 |
+
|
21 |
+
from eva_clip import create_model_and_transforms
|
22 |
+
from eva_clip.constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
|
23 |
+
from toonmage.encoders import IDEncoder
|
24 |
+
from toonmage.utils import is_torch2_available
|
25 |
+
|
26 |
+
if is_torch2_available():
|
27 |
+
from toonmage.attention_processor import AttnProcessor2_0 as AttnProcessor
|
28 |
+
from toonmage.attention_processor import IDAttnProcessor2_0 as IDAttnProcessor
|
29 |
+
else:
|
30 |
+
from toonmage.attention_processor import AttnProcessor, IDAttnProcessor
|
31 |
+
|
32 |
+
|
33 |
+
class ToonMagePipeline:
|
34 |
+
def __init__(self, *args, **kwargs):
|
35 |
+
super().__init__()
|
36 |
+
self.device = 'cuda'
|
37 |
+
sdxl_base_repo = 'stabilityai/stable-diffusion-xl-base-1.0'
|
38 |
+
sdxl_lightning_repo = 'ByteDance/SDXL-Lightning'
|
39 |
+
self.sdxl_base_repo = sdxl_base_repo
|
40 |
+
|
41 |
+
# load base model
|
42 |
+
unet = UNet2DConditionModel.from_config(sdxl_base_repo, subfolder='unet').to(self.device, torch.float16)
|
43 |
+
unet.load_state_dict(
|
44 |
+
load_file(
|
45 |
+
hf_hub_download(sdxl_lightning_repo, 'sdxl_lightning_4step_unet.safetensors'), device=self.device
|
46 |
+
)
|
47 |
+
)
|
48 |
+
unet.half()
|
49 |
+
self.hack_unet_attn_layers(unet)
|
50 |
+
self.pipe = StableDiffusionXLPipeline.from_pretrained(
|
51 |
+
sdxl_base_repo, unet=unet, torch_dtype=torch.float16, variant="fp16"
|
52 |
+
).to(self.device)
|
53 |
+
self.pipe.watermark = None
|
54 |
+
|
55 |
+
# scheduler
|
56 |
+
self.pipe.scheduler = DPMSolverMultistepScheduler.from_config(
|
57 |
+
self.pipe.scheduler.config, timestep_spacing="trailing"
|
58 |
+
)
|
59 |
+
|
60 |
+
# ID adapters
|
61 |
+
self.id_adapter = IDEncoder().to(self.device)
|
62 |
+
|
63 |
+
# preprocessors
|
64 |
+
# face align and parsing
|
65 |
+
self.face_helper = FaceRestoreHelper(
|
66 |
+
upscale_factor=1,
|
67 |
+
face_size=512,
|
68 |
+
crop_ratio=(1, 1),
|
69 |
+
det_model='retinaface_resnet50',
|
70 |
+
save_ext='png',
|
71 |
+
device=self.device,
|
72 |
+
)
|
73 |
+
self.face_helper.face_parse = None
|
74 |
+
self.face_helper.face_parse = init_parsing_model(model_name='bisenet', device=self.device)
|
75 |
+
# clip-vit backbone
|
76 |
+
model, _, _ = create_model_and_transforms('EVA02-CLIP-L-14-336', 'eva_clip', force_custom_clip=True)
|
77 |
+
model = model.visual
|
78 |
+
self.clip_vision_model = model.to(self.device)
|
79 |
+
eva_transform_mean = getattr(self.clip_vision_model, 'image_mean', OPENAI_DATASET_MEAN)
|
80 |
+
eva_transform_std = getattr(self.clip_vision_model, 'image_std', OPENAI_DATASET_STD)
|
81 |
+
if not isinstance(eva_transform_mean, (list, tuple)):
|
82 |
+
eva_transform_mean = (eva_transform_mean,) * 3
|
83 |
+
if not isinstance(eva_transform_std, (list, tuple)):
|
84 |
+
eva_transform_std = (eva_transform_std,) * 3
|
85 |
+
self.eva_transform_mean = eva_transform_mean
|
86 |
+
self.eva_transform_std = eva_transform_std
|
87 |
+
# antelopev2
|
88 |
+
snapshot_download('DIAMONIK7777/antelopev2', local_dir='models/antelopev2')
|
89 |
+
self.app = FaceAnalysis(
|
90 |
+
name='antelopev2', root='.', providers=['CPUExecutionProvider']
|
91 |
+
)
|
92 |
+
self.app.prepare(ctx_id=0, det_size=(640, 640))
|
93 |
+
self.handler_ante = insightface.model_zoo.get_model('models/antelopev2/glintr100.onnx', providers=['CPUExecutionProvider'])
|
94 |
+
self.handler_ante.prepare(ctx_id=0)
|
95 |
+
|
96 |
+
print('load done')
|
97 |
+
|
98 |
+
gc.collect()
|
99 |
+
torch.cuda.empty_cache()
|
100 |
+
|
101 |
+
self.load_pretrain()
|
102 |
+
|
103 |
+
# other configs
|
104 |
+
self.debug_img_list = []
|
105 |
+
|
106 |
+
def hack_unet_attn_layers(self, unet):
|
107 |
+
id_adapter_attn_procs = {}
|
108 |
+
for name, _ in unet.attn_processors.items():
|
109 |
+
cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
|
110 |
+
if name.startswith("mid_block"):
|
111 |
+
hidden_size = unet.config.block_out_channels[-1]
|
112 |
+
elif name.startswith("up_blocks"):
|
113 |
+
block_id = int(name[len("up_blocks.")])
|
114 |
+
hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
|
115 |
+
elif name.startswith("down_blocks"):
|
116 |
+
block_id = int(name[len("down_blocks.")])
|
117 |
+
hidden_size = unet.config.block_out_channels[block_id]
|
118 |
+
if cross_attention_dim is not None:
|
119 |
+
id_adapter_attn_procs[name] = IDAttnProcessor(
|
120 |
+
hidden_size=hidden_size,
|
121 |
+
cross_attention_dim=cross_attention_dim,
|
122 |
+
).to(unet.device)
|
123 |
+
else:
|
124 |
+
id_adapter_attn_procs[name] = AttnProcessor()
|
125 |
+
unet.set_attn_processor(id_adapter_attn_procs)
|
126 |
+
self.id_adapter_attn_layers = nn.ModuleList(unet.attn_processors.values())
|
127 |
+
|
128 |
+
def load_pretrain(self):
|
129 |
+
hf_hub_download('SunderAli17/SAK', 'toonmage_v2.bin', local_dir='models')
|
130 |
+
ckpt_path = 'models/toonmage_v2.bin'
|
131 |
+
state_dict = torch.load(ckpt_path, map_location='cpu')
|
132 |
+
state_dict_dict = {}
|
133 |
+
for k, v in state_dict.items():
|
134 |
+
module = k.split('.')[0]
|
135 |
+
state_dict_dict.setdefault(module, {})
|
136 |
+
new_k = k[len(module) + 1 :]
|
137 |
+
state_dict_dict[module][new_k] = v
|
138 |
+
|
139 |
+
for module in state_dict_dict:
|
140 |
+
print(f'loading from {module}')
|
141 |
+
getattr(self, module).load_state_dict(state_dict_dict[module], strict=True)
|
142 |
+
|
143 |
+
def to_gray(self, img):
|
144 |
+
x = 0.299 * img[:, 0:1] + 0.587 * img[:, 1:2] + 0.114 * img[:, 2:3]
|
145 |
+
x = x.repeat(1, 3, 1, 1)
|
146 |
+
return x
|
147 |
+
|
148 |
+
def get_id_embedding(self, image):
|
149 |
+
"""
|
150 |
+
Args:
|
151 |
+
image: numpy rgb image, range [0, 255]
|
152 |
+
"""
|
153 |
+
self.face_helper.clean_all()
|
154 |
+
image_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
|
155 |
+
# get antelopev2 embedding
|
156 |
+
face_info = self.app.get(image_bgr)
|
157 |
+
if len(face_info) > 0:
|
158 |
+
face_info = sorted(face_info, key=lambda x: (x['bbox'][2] - x['bbox'][0]) * x['bbox'][3] - x['bbox'][1])[
|
159 |
+
-1
|
160 |
+
] # only use the maximum face
|
161 |
+
id_ante_embedding = face_info['embedding']
|
162 |
+
self.debug_img_list.append(
|
163 |
+
image[
|
164 |
+
int(face_info['bbox'][1]) : int(face_info['bbox'][3]),
|
165 |
+
int(face_info['bbox'][0]) : int(face_info['bbox'][2]),
|
166 |
+
]
|
167 |
+
)
|
168 |
+
else:
|
169 |
+
id_ante_embedding = None
|
170 |
+
|
171 |
+
# using facexlib to detect and align face
|
172 |
+
self.face_helper.read_image(image_bgr)
|
173 |
+
self.face_helper.get_face_landmarks_5(only_center_face=True)
|
174 |
+
self.face_helper.align_warp_face()
|
175 |
+
if len(self.face_helper.cropped_faces) == 0:
|
176 |
+
raise RuntimeError('facexlib align face fail')
|
177 |
+
align_face = self.face_helper.cropped_faces[0]
|
178 |
+
# incase insightface didn't detect face
|
179 |
+
if id_ante_embedding is None:
|
180 |
+
print('fail to detect face using insightface, extract embedding on align face')
|
181 |
+
id_ante_embedding = self.handler_ante.get_feat(align_face)
|
182 |
+
|
183 |
+
id_ante_embedding = torch.from_numpy(id_ante_embedding).to(self.device)
|
184 |
+
if id_ante_embedding.ndim == 1:
|
185 |
+
id_ante_embedding = id_ante_embedding.unsqueeze(0)
|
186 |
+
|
187 |
+
# parsing
|
188 |
+
input = img2tensor(align_face, bgr2rgb=True).unsqueeze(0) / 255.0
|
189 |
+
input = input.to(self.device)
|
190 |
+
parsing_out = self.face_helper.face_parse(normalize(input, [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]))[0]
|
191 |
+
parsing_out = parsing_out.argmax(dim=1, keepdim=True)
|
192 |
+
bg_label = [0, 16, 18, 7, 8, 9, 14, 15]
|
193 |
+
bg = sum(parsing_out == i for i in bg_label).bool()
|
194 |
+
white_image = torch.ones_like(input)
|
195 |
+
# only keep the face features
|
196 |
+
face_features_image = torch.where(bg, white_image, self.to_gray(input))
|
197 |
+
self.debug_img_list.append(tensor2img(face_features_image, rgb2bgr=False))
|
198 |
+
|
199 |
+
# transform img before sending to eva-clip-vit
|
200 |
+
face_features_image = resize(face_features_image, self.clip_vision_model.image_size, InterpolationMode.BICUBIC)
|
201 |
+
face_features_image = normalize(face_features_image, self.eva_transform_mean, self.eva_transform_std)
|
202 |
+
id_cond_vit, id_vit_hidden = self.clip_vision_model(
|
203 |
+
face_features_image, return_all_features=False, return_hidden=True, shuffle=False
|
204 |
+
)
|
205 |
+
id_cond_vit_norm = torch.norm(id_cond_vit, 2, 1, True)
|
206 |
+
id_cond_vit = torch.div(id_cond_vit, id_cond_vit_norm)
|
207 |
+
|
208 |
+
id_cond = torch.cat([id_ante_embedding, id_cond_vit], dim=-1)
|
209 |
+
id_uncond = torch.zeros_like(id_cond)
|
210 |
+
id_vit_hidden_uncond = []
|
211 |
+
for layer_idx in range(0, len(id_vit_hidden)):
|
212 |
+
id_vit_hidden_uncond.append(torch.zeros_like(id_vit_hidden[layer_idx]))
|
213 |
+
|
214 |
+
id_embedding = self.id_adapter(id_cond, id_vit_hidden)
|
215 |
+
uncond_id_embedding = self.id_adapter(id_uncond, id_vit_hidden_uncond)
|
216 |
+
|
217 |
+
# return id_embedding
|
218 |
+
return torch.cat((uncond_id_embedding, id_embedding), dim=0)
|
219 |
+
|
220 |
+
def inference(self, prompt, size, prompt_n='', image_embedding=None, id_scale=1.0, guidance_scale=1.2, steps=4):
|
221 |
+
images = self.pipe(
|
222 |
+
prompt=prompt,
|
223 |
+
negative_prompt=prompt_n,
|
224 |
+
num_images_per_prompt=size[0],
|
225 |
+
height=size[1],
|
226 |
+
width=size[2],
|
227 |
+
num_inference_steps=steps,
|
228 |
+
guidance_scale=guidance_scale,
|
229 |
+
cross_attention_kwargs={'id_embedding': image_embedding, 'id_scale': id_scale},
|
230 |
+
).images
|
231 |
+
|
232 |
+
return images
|
toonmage/utils.py
ADDED
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import importlib
|
2 |
+
import os
|
3 |
+
import random
|
4 |
+
|
5 |
+
import cv2
|
6 |
+
import numpy as np
|
7 |
+
import torch
|
8 |
+
import torch.nn.functional as F
|
9 |
+
from transformers import PretrainedConfig
|
10 |
+
|
11 |
+
|
12 |
+
def seed_everything(seed):
|
13 |
+
os.environ["PL_GLOBAL_SEED"] = str(seed)
|
14 |
+
random.seed(seed)
|
15 |
+
np.random.seed(seed)
|
16 |
+
torch.manual_seed(seed)
|
17 |
+
torch.cuda.manual_seed_all(seed)
|
18 |
+
|
19 |
+
|
20 |
+
def is_torch2_available():
|
21 |
+
return hasattr(F, "scaled_dot_product_attention")
|
22 |
+
|
23 |
+
|
24 |
+
def instantiate_from_config(config):
|
25 |
+
if "target" not in config:
|
26 |
+
if config == '__is_first_stage__' or config == "__is_unconditional__":
|
27 |
+
return None
|
28 |
+
raise KeyError("Expected key `target` to instantiate.")
|
29 |
+
return get_obj_from_str(config["target"])(**config.get("params", {}))
|
30 |
+
|
31 |
+
|
32 |
+
def get_obj_from_str(string, reload=False):
|
33 |
+
module, cls = string.rsplit(".", 1)
|
34 |
+
if reload:
|
35 |
+
module_imp = importlib.import_module(module)
|
36 |
+
importlib.reload(module_imp)
|
37 |
+
return getattr(importlib.import_module(module, package=None), cls)
|
38 |
+
|
39 |
+
|
40 |
+
def drop_seq_token(seq, drop_rate=0.5):
|
41 |
+
idx = torch.randperm(seq.size(1))
|
42 |
+
num_keep_tokens = int(len(idx) * (1 - drop_rate))
|
43 |
+
idx = idx[:num_keep_tokens]
|
44 |
+
seq = seq[:, idx]
|
45 |
+
return seq
|
46 |
+
|
47 |
+
|
48 |
+
def import_model_class_from_model_name_or_path(
|
49 |
+
pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
|
50 |
+
):
|
51 |
+
text_encoder_config = PretrainedConfig.from_pretrained(
|
52 |
+
pretrained_model_name_or_path, subfolder=subfolder, revision=revision
|
53 |
+
)
|
54 |
+
model_class = text_encoder_config.architectures[0]
|
55 |
+
|
56 |
+
if model_class == "CLIPTextModel":
|
57 |
+
from transformers import CLIPTextModel
|
58 |
+
|
59 |
+
return CLIPTextModel
|
60 |
+
elif model_class == "CLIPTextModelWithProjection": # noqa RET505
|
61 |
+
from transformers import CLIPTextModelWithProjection
|
62 |
+
|
63 |
+
return CLIPTextModelWithProjection
|
64 |
+
else:
|
65 |
+
raise ValueError(f"{model_class} is not supported.")
|
66 |
+
|
67 |
+
|
68 |
+
def resize_numpy_image_long(image, resize_long_edge=768):
|
69 |
+
h, w = image.shape[:2]
|
70 |
+
if max(h, w) <= resize_long_edge:
|
71 |
+
return image
|
72 |
+
k = resize_long_edge / max(h, w)
|
73 |
+
h = int(h * k)
|
74 |
+
w = int(w * k)
|
75 |
+
image = cv2.resize(image, (w, h), interpolation=cv2.INTER_LANCZOS4)
|
76 |
+
return image
|