Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,022 Bytes
323d67d |
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 |
import gc
import cv2
import insightface
import numpy as np
import torch
import torch.nn as nn
from pulid.utils import img2tensor, tensor2img
from diffusers import DPMSolverMultistepScheduler, StableDiffusionXLPipeline
from facexlib.parsing import init_parsing_model
from facexlib.utils.face_restoration_helper import FaceRestoreHelper
from huggingface_hub import hf_hub_download, snapshot_download
from insightface.app import FaceAnalysis
from safetensors.torch import load_file
from torchvision.transforms import InterpolationMode
from torchvision.transforms.functional import normalize, resize
from eva_clip import create_model_and_transforms
from eva_clip.constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
from pulid.encoders_transformer import IDFormer
from pulid.utils import is_torch2_available, sample_dpmpp_2m, sample_dpmpp_sde
if is_torch2_available():
from pulid.attention_processor import AttnProcessor2_0 as AttnProcessor
from pulid.attention_processor import IDAttnProcessor2_0 as IDAttnProcessor
else:
from pulid.attention_processor import AttnProcessor, IDAttnProcessor
class PuLIDEncoder:
def __init__(
self,
device
):
super().__init__()
self.device = device
# scheduler
# self.pipe.scheduler = DPMSolverMultistepScheduler.from_config(
# self.pipe.scheduler.config
# )
# ID adapters
# self.id_adapter = IDFormer().to(self.device)
# preprocessors
# face align and parsing
self.face_helper = FaceRestoreHelper(
upscale_factor=1,
face_size=512,
crop_ratio=(1, 1),
det_model="retinaface_resnet50",
save_ext="png",
device=self.device,
)
self.face_helper.face_parse = None
self.face_helper.face_parse = init_parsing_model(
model_name="bisenet", device=self.device
)
# clip-vit backbone
model, _, _ = create_model_and_transforms(
"EVA02-CLIP-L-14-336", "eva_clip", force_custom_clip=True
)
model = model.visual
self.clip_vision_model = model.to(self.device)
eva_transform_mean = getattr(
self.clip_vision_model, "image_mean", OPENAI_DATASET_MEAN
)
eva_transform_std = getattr(
self.clip_vision_model, "image_std", OPENAI_DATASET_STD
)
if not isinstance(eva_transform_mean, (list, tuple)):
eva_transform_mean = (eva_transform_mean,) * 3
if not isinstance(eva_transform_std, (list, tuple)):
eva_transform_std = (eva_transform_std,) * 3
self.eva_transform_mean = eva_transform_mean
self.eva_transform_std = eva_transform_std
# antelopev2
snapshot_download("DIAMONIK7777/antelopev2", local_dir="models/antelopev2")
self.app = FaceAnalysis(
name="antelopev2",
root=".",
providers=["CPUExecutionProvider"],
)
self.app.prepare(ctx_id=0, det_size=(640, 640))
self.handler_ante = insightface.model_zoo.get_model(
"models/antelopev2/glintr100.onnx"
)
self.handler_ante.prepare(ctx_id=0)
gc.collect()
torch.cuda.empty_cache()
# self.load_pretrain()
# other configs
self.debug_img_list = []
def to_gray(self, img):
x = 0.299 * img[:, 0:1] + 0.587 * img[:, 1:2] + 0.114 * img[:, 2:3]
x = x.repeat(1, 3, 1, 1)
return x
def get_id_embedding(self, image_list):
"""
Args:
image in image_list: numpy rgb image, range [0, 255]
"""
id_cond_list = []
id_vit_hidden_list = []
for ii, image in enumerate(image_list):
self.face_helper.clean_all()
image_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
# get antelopev2 embedding
face_info = self.app.get(image_bgr)
if len(face_info) > 0:
face_info = sorted(
face_info,
key=lambda x: (x["bbox"][2] - x["bbox"][0])
* (x["bbox"][3] - x["bbox"][1]),
)[
-1
] # only use the maximum face
id_ante_embedding = face_info["embedding"]
self.debug_img_list.append(
image[
int(face_info["bbox"][1]) : int(face_info["bbox"][3]),
int(face_info["bbox"][0]) : int(face_info["bbox"][2]),
]
)
else:
id_ante_embedding = None
# using facexlib to detect and align face
self.face_helper.read_image(image_bgr)
self.face_helper.get_face_landmarks_5(only_center_face=True)
self.face_helper.align_warp_face()
if len(self.face_helper.cropped_faces) == 0:
raise RuntimeError("facexlib align face fail")
align_face = self.face_helper.cropped_faces[0]
# incase insightface didn't detect face
if id_ante_embedding is None:
print(
"fail to detect face using insightface, extract embedding on align face"
)
id_ante_embedding = self.handler_ante.get_feat(align_face)
id_ante_embedding = torch.from_numpy(id_ante_embedding).to(self.device)
if id_ante_embedding.ndim == 1:
id_ante_embedding = id_ante_embedding.unsqueeze(0)
# parsing
input = img2tensor(align_face, bgr2rgb=True).unsqueeze(0) / 255.0
input = input.to(self.device)
parsing_out = self.face_helper.face_parse(
normalize(input, [0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
)[0]
parsing_out = parsing_out.argmax(dim=1, keepdim=True)
bg_label = [0, 16, 18, 7, 8, 9, 14, 15]
bg = sum(parsing_out == i for i in bg_label).bool()
white_image = torch.ones_like(input)
# only keep the face features
face_features_image = torch.where(bg, white_image, self.to_gray(input))
self.debug_img_list.append(tensor2img(face_features_image, rgb2bgr=False))
# transform img before sending to eva-clip-vit
face_features_image = resize(
face_features_image,
self.clip_vision_model.image_size,
InterpolationMode.BICUBIC,
)
face_features_image = normalize(
face_features_image, self.eva_transform_mean, self.eva_transform_std
)
id_cond_vit, id_vit_hidden = self.clip_vision_model(
face_features_image,
return_all_features=False,
return_hidden=True,
shuffle=False,
)
id_cond_vit_norm = torch.norm(id_cond_vit, 2, 1, True)
id_cond_vit = torch.div(id_cond_vit, id_cond_vit_norm)
id_cond = torch.cat([id_ante_embedding, id_cond_vit], dim=-1)
id_cond_list.append(id_cond)
id_vit_hidden_list.append(id_vit_hidden)
id_uncond = torch.zeros_like(id_cond_list[0])
id_vit_hidden_uncond = []
for layer_idx in range(0, len(id_vit_hidden_list[0])):
id_vit_hidden_uncond.append(
torch.zeros_like(id_vit_hidden_list[0][layer_idx])
)
id_cond = torch.stack(id_cond_list, dim=1)
id_vit_hidden = id_vit_hidden_list[0]
for i in range(1, len(image_list)):
for j, x in enumerate(id_vit_hidden_list[i]):
id_vit_hidden[j] = torch.cat([id_vit_hidden[j], x], dim=1)
# id_embedding = self.id_adapter(id_cond, id_vit_hidden)
# uncond_id_embedding = self.id_adapter(id_uncond, id_vit_hidden_uncond)
# return id_embedding
return id_cond, id_vit_hidden, id_uncond, id_vit_hidden_uncond |