Spaces:
Running on Zero
Running on Zero
File size: 8,171 Bytes
a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b a891a57 915d52b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 | # coding: utf-8
"""
Pipeline for gradio
已适配 Hugging Face ZeroGPU:
1. 给会使用 CUDA 的 Gradio 回调函数添加 @spaces.GPU 装饰器。
2. 保留 Gradio 4.x 兼容写法。
3. 如果本地环境没有 spaces 包,也不会影响本地普通运行。
"""
import gradio as gr
# Hugging Face ZeroGPU 必须使用 spaces.GPU 装饰需要 GPU 的函数。
# 为了兼容本地运行,如果没有安装 spaces,则使用一个空装饰器。
try:
import spaces
except Exception:
class _DummySpaces:
@staticmethod
def GPU(duration=120):
def decorator(func):
return func
return decorator
spaces = _DummySpaces()
from .config.argument_config import ArgumentConfig
from .live_portrait_pipeline import LivePortraitPipeline
from .utils.io import load_img_online
from .utils.rprint import rlog as log
from .utils.crop import prepare_paste_back, paste_back
from .utils.camera import get_rotation_matrix
from .utils.retargeting_utils import calc_eye_close_ratio, calc_lip_close_ratio
def update_args(args, user_args):
"""
根据用户输入更新参数。
"""
for k, v in user_args.items():
if hasattr(args, k):
setattr(args, k, v)
return args
class GradioPipeline(LivePortraitPipeline):
def __init__(self, inference_cfg, crop_cfg, args: ArgumentConfig):
super().__init__(inference_cfg, crop_cfg)
self.args = args
# 单图重定向状态缓存
self.start_prepare = False
self.f_s_user = None
self.x_c_s_info_user = None
self.x_s_user = None
self.source_lmk_user = None
self.mask_ori = None
self.img_rgb = None
self.crop_M_c2o = None
self.I_s_vis = None
@spaces.GPU(duration=300)
def execute_video(
self,
input_image_path,
input_video_path,
flag_relative_input,
flag_do_crop_input,
flag_remap_input,
):
"""
视频驱动肖像动画。
注意:
这个函数内部会调用模型推理和 CUDA,因此必须放在 @spaces.GPU 里。
"""
if input_image_path is not None and input_video_path is not None:
args_user = {
"source_image": input_image_path,
"driving_info": input_video_path,
"flag_relative": flag_relative_input,
"flag_do_crop": flag_do_crop_input,
"flag_pasteback": flag_remap_input,
}
# 根据用户输入更新配置
self.args = update_args(self.args, args_user)
self.live_portrait_wrapper.update_config(self.args.__dict__)
self.cropper.update_config(self.args.__dict__)
# 执行视频驱动动画
video_path, video_path_concat = self.execute(self.args)
return video_path, video_path_concat
raise gr.Error(
"The input source portrait or driving video hasn't been prepared yet 💥!",
duration=5,
)
@spaces.GPU(duration=180)
def execute_image(self, input_eye_ratio: float, input_lip_ratio: float):
"""
单图表情重定向。
注意:
这里会执行 .to("cuda")、retarget_eye、retarget_lip、warp_decode,
因此必须放在 @spaces.GPU 里。
"""
if input_eye_ratio is None or input_lip_ratio is None:
raise gr.Error("Invalid ratio input 💥!", duration=5)
if self.f_s_user is None:
if self.start_prepare:
raise gr.Error(
"The source portrait is under processing 💥! Please wait for a second.",
duration=5,
)
raise gr.Error(
"The source portrait hasn't been prepared yet 💥! Please scroll to the top of the page to upload.",
duration=5,
)
x_s_user = self.x_s_user.to("cuda")
f_s_user = self.f_s_user.to("cuda")
# 计算眼睛重定向
combined_eye_ratio_tensor = self.live_portrait_wrapper.calc_combined_eye_ratio(
[[input_eye_ratio]],
self.source_lmk_user,
)
eyes_delta = self.live_portrait_wrapper.retarget_eye(
x_s_user,
combined_eye_ratio_tensor,
)
# 计算嘴唇重定向
combined_lip_ratio_tensor = self.live_portrait_wrapper.calc_combined_lip_ratio(
[[input_lip_ratio]],
self.source_lmk_user,
)
lip_delta = self.live_portrait_wrapper.retarget_lip(
x_s_user,
combined_lip_ratio_tensor,
)
num_kp = x_s_user.shape[1]
# 默认基于 x_s 做变形
x_d_new = (
x_s_user
+ eyes_delta.reshape(-1, num_kp, 3)
+ lip_delta.reshape(-1, num_kp, 3)
)
# 解码输出
out = self.live_portrait_wrapper.warp_decode(
f_s_user,
x_s_user,
x_d_new,
)
out = self.live_portrait_wrapper.parse_output(out["out"])[0]
# 贴回原图
out_to_ori_blend = paste_back(
out,
self.crop_M_c2o,
self.img_rgb,
self.mask_ori,
)
return out, out_to_ori_blend
@spaces.GPU(duration=180)
def prepare_retargeting(self, input_image_path, flag_do_crop=True):
"""
单图表情重定向的预处理。
注意:
日志中的报错发生在这个函数调用链里:
prepare_retargeting -> prepare_source -> x.cuda(...)
所以这个函数必须使用 @spaces.GPU。
"""
if input_image_path is not None:
self.start_prepare = True
inference_cfg = self.live_portrait_wrapper.cfg
# 读取源图
img_rgb = load_img_online(
input_image_path,
mode="rgb",
max_dim=1280,
n=16,
)
log(f"Load source image from {input_image_path}.")
# 裁剪人脸
crop_info = self.cropper.crop_single_image(img_rgb)
if flag_do_crop:
I_s = self.live_portrait_wrapper.prepare_source(
crop_info["img_crop_256x256"]
)
else:
I_s = self.live_portrait_wrapper.prepare_source(img_rgb)
# 提取关键点信息
x_s_info = self.live_portrait_wrapper.get_kp_info(I_s)
# 保留原逻辑:计算旋转矩阵
# 当前变量暂未在后续使用,但保留,避免影响原项目行为。
_ = get_rotation_matrix(
x_s_info["pitch"],
x_s_info["yaw"],
x_s_info["roll"],
)
# 缓存后续单图重定向需要的数据
self.f_s_user = self.live_portrait_wrapper.extract_feature_3d(I_s)
self.x_s_user = self.live_portrait_wrapper.transform_keypoint(x_s_info)
self.x_s_info_user = x_s_info
self.source_lmk_user = crop_info["lmk_crop"]
self.img_rgb = img_rgb
self.crop_M_c2o = crop_info["M_c2o"]
self.mask_ori = prepare_paste_back(
inference_cfg.mask_crop,
crop_info["M_c2o"],
dsize=(img_rgb.shape[1], img_rgb.shape[0]),
)
# 更新滑块默认值
eye_close_ratio = calc_eye_close_ratio(self.source_lmk_user[None])
eye_close_ratio = float(eye_close_ratio.squeeze(0).mean())
lip_close_ratio = calc_lip_close_ratio(self.source_lmk_user[None])
lip_close_ratio = float(lip_close_ratio.squeeze(0).mean())
# 预览图
self.I_s_vis = self.live_portrait_wrapper.parse_output(I_s)[0]
self.start_prepare = False
return eye_close_ratio, lip_close_ratio, self.I_s_vis
# 点击清空按钮时走这里
if self.I_s_vis is not None:
return 0.8, 0.8, self.I_s_vis
return 0.8, 0.8, None |