roop: one click face swap. 只用一張人臉圖片，就能完成視頻換臉。
- 項(xiàng)目地址： https://github.com/s0md3v/roop
- Colab 部署： https://github.com/dream80/roop_colab
本文是本地部署的實(shí)踐記錄。
環(huán)境基礎(chǔ)
- OS: Ubuntu 22.04.2 LTS
- Kernel: 5.19.0
- CPU: Intel(R) Core(TM) i7-10750H
- GPU: NVIDIA GeForce RTX 2060
- RAM: 16.0 GiB
環(huán)境安裝
Nvidia 驅(qū)動(dòng)
# 查看可用驅(qū)動(dòng)信息
ubuntu-drivers devices
# 安裝 Nvidia 驅(qū)動(dòng)（可選提示 recommended 的）
sudo apt install nvidia-driver-530
# 查看 Nvidia 驅(qū)動(dòng)（查看不了，可重登錄或重啟）
nvidia-smi
Anaconda
- Anaconda: https://www.anaconda.com/download
# 安裝 Anaconda, 都 yes
bash Anaconda3-2023.07-1-Linux-x86_64.sh
創(chuàng)建虛擬環(huán)境，
conda create -n roop python=3.10 -y
conda activate roop
roop 安裝
獲取代碼，
sudo apt install git -y
git clone --depth 1 https://github.com/s0md3v/roop.git
安裝依賴，
cd roop
pip install -r requirements.txt
# 如遇錯(cuò)誤，可依提示。例如：
# pip install -r requirements.txt --use-pep517
# 如遇 dependency conflict，可修改依賴版本。例如：
# numpy>=1.23.5
# 若配置鏡像
pip config set global.index-url http://mirrors.aliyun.com/pypi/simple
pip config set install.trusted-host mirrors.aliyun.com
cat ~/.config/pip/pip.conf
roop 體驗(yàn)
運(yùn)行，
cd roop
python run.py --execution-provider cuda
# 如果報(bào)錯(cuò)類似 ONNXRuntimeError CUBLAS_STATUS_ALLOC_FAILED，需減少執(zhí)行線程
# https://github.com/s0md3v/roop/issues/413
python run.py --execution-provider cuda --execution-threads 1
可見窗口，
選擇一張人臉圖片、一個(gè)視頻，就能完成轉(zhuǎn)換了。
輸出信息，
$ python run.py --execution-provider cuda --execution-threads 1
Applied providers: ['CUDAExecutionProvider', 'CPUExecutionProvider'], with options: {'CPUExecutionProvider': {}, 'CUDAExecutionProvider': {'device_id': '0', 'gpu_mem_limit': '18446744073709551615', 'gpu_external_alloc': '0', 'gpu_external_free': '0', 'gpu_external_empty_cache': '0', 'cudnn_conv_algo_search': 'EXHAUSTIVE', 'cudnn_conv1d_pad_to_nc1d': '0', 'arena_extend_strategy': 'kNextPowerOfTwo', 'do_copy_in_default_stream': '1', 'enable_cuda_graph': '0', 'cudnn_conv_use_max_workspace': '1', 'tunable_op_enable': '0', 'enable_skip_layer_norm_strict_mode': '0', 'tunable_op_tuning_enable': '0'}}
# ...
find model: /home/john/.insightface/models/buffalo_l/w600k_r50.onnx recognition ['None', 3, 112, 112] 127.5 127.5
set det-size: (640, 640)
inswapper-shape: [1, 3, 128, 128]
100%|█████████████████████████████████████████████████████████████████████████████████████████| 336/336 [00:01<00:00, 185.49it/s]
[ROOP.CORE] Creating temp resources...
[ROOP.CORE] Extracting frames with 30.0 FPS...
[ROOP.FACE-SWAPPER] Progressing...
Processing: 100%|█| 336/336 [01:40<00:00, 3.33frame/s, memory_usage=04.94GB, execution_providers=['CUDAExecutionProvider'], exec
[ROOP.CORE] Creating video with 30.0 FPS...
[ROOP.CORE] Restoring audio...
[ROOP.CORE] Processing to video succeed!
結(jié)語
roop 視頻換臉，個(gè)人感覺效果還不錯(cuò)。
GoCoding 個(gè)人實(shí)踐的經(jīng)驗(yàn)分享，歡迎關(guān)注！
