Faceswap 3 open for testing

[teskki]

Hi I need help with this installation process... I tried installing the Faceswap 3 on my kali linux and this is what am getting

(base) ┌──(tejiri㉿kali)-[~]
└─$ git clone https://github.com/deepfakes/faceswap.git fs3
Cloning into 'fs3'...
remote: Enumerating objects: 14675, done.
remote: Counting objects: 100% (1062/1062), done.
remote: Compressing objects: 100% (624/624), done.
error: RPC failed; curl 18 transfer closed with outstanding read data remaining
error: 4480 bytes of body are still expected
fetch-pack: unexpected disconnect while reading sideband packet
fatal: early EOF
fatal: fetch-pack: invalid index-pack output

[torzdf]

@ianstephens Thanks for testing. Unfortunately I have been very busy of late, so not much time for development.

I have made a note of your feedback and will investigate as soon as I get an opportunity.

[torzdf]

Hi I need help with this installation process... I tried installing the Faceswap 3 on my kali linux and this is what am getting

(base) ┌──(tejiri㉿kali)-[~]
└─$ git clone https://github.com/deepfakes/faceswap.git fs3
Cloning into 'fs3'...
remote: Enumerating objects: 14675, done.
remote: Counting objects: 100% (1062/1062), done.
remote: Compressing objects: 100% (624/624), done.
error: RPC failed; curl 18 transfer closed with outstanding read data remaining
error: 4480 bytes of body are still expected
fetch-pack: unexpected disconnect while reading sideband packet
fatal: early EOF
fatal: fetch-pack: invalid index-pack output

This is a network connection issue. It looks like your ISP is cutting off your connection before the download completes.

You can try either:

• Using an internet connection elsewhere
• Trying to enable a VPN (sometimes works for this issue)

[Ryzen1988]

Hey wow, am i late to the party, damn. But it gives me this strange error because torchvision is already installed. I added crash report
All other masks besides BiSeNet seem to work perfect

Nevermind, fixed it by uninstalling and reinstalling different toch, now version 2.3.0

Initialized: BiSeNet
ImportError: Exception encountered when calling UpSampling2D.call().

[1mThe torchvision package is necessary to use resize with the torch backend. Please install torchvision.[0m

Arguments received by UpSampling2D.call():
• inputs=torch.Tensor(shape=torch.Size([8, 32, 32, 5]), dtype=float32)

On a more cheerful note, the manual tool feels a lot quicker and a bit more robuust. really looking forward to try out the new convnext and the new AdamW optimizer.

[Ryzen1988]

Have some errors that appear when trying to train a model

Code: Select all

C:\Users\Admin\.conda\envs\fs3\Lib\site-packages\keras\src\backend\torch\nn.py:416: UserWarning: Plan failed with a cudnnException: CUDNN_BACKEND_EXECUTION_PLAN_DESCRIPTOR: cudnnFinalize Descriptor Failed cudnn_status: CUDNN_STATUS_NOT_SUPPORTED (Triggered internally at ..\aten\src\ATen\native\cudnn\Conv_v8.cpp:919.)
outputs = tnn.conv2d(
C:\Users\Admin\.conda\envs\fs3\Lib\site-packages\keras\src\backend\torch\nn.py:416: UserWarning: Using padding='same' with even kernel lengths and odd dilation may require a zero-padded copy of the input be created (Triggered internally at ..\aten\src\ATen\native\Convolution.cpp:1032.)
outputs = tnn.conv2d(
C:\Users\Admin\.conda\envs\fs3\Lib\site-packages\torch\autograd\graph.py:744: UserWarning: Plan failed with a cudnnException: CUDNN_BACKEND_EXECUTION_PLAN_DESCRIPTOR: cudnnFinalize Descriptor Failed cudnn_status: CUDNN_STATUS_NOT_SUPPORTED (Triggered internally at ..\aten\src\ATen\native\cudnn\Conv_v8.cpp:919.)
return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass

[Playstation5fanatic]

Hey wow, am i late to the party, damn. But it gives me this strange error because torchvision is already installed. I added crash report
All other masks besides BiSeNet seem to work perfect

Nevermind, fixed it by uninstalling and reinstalling different toch, now version 2.3.0

Initialized: BiSeNet
ImportError: Exception encountered when calling UpSampling2D.call().

[1mThe torchvision package is necessary to use resize with the torch backend. Please install torchvision.[0m

Arguments received by UpSampling2D.call():
• inputs=torch.Tensor(shape=torch.Size([8, 32, 32, 5]), dtype=float32)

On a more cheerful note, the manual tool feels a lot quicker and a bit more robuust. really looking forward to try out the new convnext and the new AdamW optimizer.

Hey! I'm having the same issue. What exactly did you do to fix it? I tried uninstall torchvision and reinstalling and older version. Same issue. Then I updated Faceswap via the gui, still having the same issue.

Any help would be appreciated!

[spewfy]

Problem with Alignments

Testing on Linux only, I have a problem with alignments file. All frames are missing alignments so convert process isn't working. Testing Alignments with tools.py, I can identify the issue:

python faceswap.py extract -i test.mp4 -o output -p test.fsa -A fan -D s3fd -L VERBOSE
01/19/2025 12:08:52 INFO Aligner filtered: (features (1.00): 7)
01/19/2025 12:08:52 INFO Writing alignments to: '/home/ubuntu/fs3/test.fsa'
01/19/2025 12:08:52 INFO -------------------------
01/19/2025 12:08:52 INFO Images found: 0
01/19/2025 12:08:52 INFO Faces detected: 601
01/19/2025 12:08:52 INFO -------------------------
01/19/2025 12:08:52 INFO Process Successfully Completed. Shutting Down...

python tools.py alignments -j extract -a test.fsa -r test.mp4 -c extract -L VERBOSE
01/19/2025 12:10:55 INFO [FRAMES DATA]
01/19/2025 12:10:55 VERBOSE Video exists at: '/home/ubuntu/fs3/test.mp4'
01/19/2025 12:10:55 INFO Loading video frames from /home/ubuntu/fs3/test.mp4
01/19/2025 12:10:56 VERBOSE 601 items loaded
01/19/2025 12:10:56 INFO [EXTRACT FACES]
01/19/2025 12:10:56 VERBOSE Creating output folder at '/home/ubuntu/fs3/extract'
01/19/2025 12:11:31 INFO 601 face(s) extracted

python tools.py alignments -j missing-alignments -a test.fsa -r extract -L VERBOSE
01/19/2025 12:13:15 INFO [FRAMES DATA]
01/19/2025 12:13:15 VERBOSE Folder exists at '/home/ubuntu/fs3/extract'
01/19/2025 12:13:15 INFO Loading file list from /home/ubuntu/fs3/extract
01/19/2025 12:13:15 VERBOSE 601 items loaded
01/19/2025 12:13:15 INFO [CHECK FRAMES]
01/19/2025 12:13:15 INFO -----------------------------------------------
01/19/2025 12:13:15 INFO --- Frames missing from alignments file (601)
01/19/2025 12:13:15 INFO -----------------------------------------------
01/19/2025 12:13:15 INFO test_000001_0.png
01/19/2025 12:13:15 INFO test_000002_0.png
01/19/2025 12:13:15 INFO test_000003_0.png
01/19/2025 12:13:15 INFO test_000004_0.png
01/19/2025 12:13:15 INFO test_000005_0.png
....

Please let me know how to get around this

Thank you

[BCD16]

Hello, M4 Max Macbook Pro user here,

Installed succesfully in a virtual miniforge3 environment, without any errors.

Training starts and after a couple seconds, this warning shows up:

/miniforge3/envs/fs3/lib/python3.11/site-packages/keras/src/backend/torch/optimizers/torch_adam.py:35: UserWarning: The operator 'aten::foreach_mul.Scalar' is not currently supported on the MPS backend and will fall back to run on the CPU. This may have performance implications. (Triggered internally at /Users/runner/work/temp/anaconda/conda-bld/pytorch_1711403251597/work/aten/src/ATen/mps/MPSFallback.mm:13.)
torch.foreach_mul_(m_list, self.beta_1)

And preview pictures like rainbow. I waited for 30 minutes, Loss is not going down. Sometimes training directly ends with NaN Protection error.

I see that your script installs pytorch 2.3.1 , i tried all versions between 2.2.2-2.3.1. Nothing changes. I changed optimizer to Lion or other ones, same result.

Same result with or without mixed precision enabled.

edit: OK now I got it working. Forced to install pytorch 2.6.0 in conda env, and override pytorch version check in /cli/launcher.py file.

Training works now with FP16 enabled, but significantly slower than tensorflow backend.

In my faceswap2 environment I have been using tensorflow 2.13 and tensorflow-metal 1.0.0 and i am really happy with its FP16 speed with legacy Adam optimizer. This beta version should be improved much much more.

[burnout_]

bug when installing for nvidia users. windows 10. username removed.

(fs3) C:\Users\\>cd fs3

(fs3) C:\Users\\fs3>git checkout fs3
branch 'fs3' set up to track 'origin/fs3'.
Switched to a new branch 'fs3'

(fs3) C:\Users\\fs3>conda activate fs3

(fs3) C:\Users\\fs3>python setup.py --installer --nvidia
←[32mINFO←[0m Running without root/admin privileges
←[32mINFO←[0m The tool provides tips for installation and installs required python packages
←[32mINFO←[0m Setup in Windows 10
←[32mINFO←[0m Installed Python: 3.11.11 64bit
←[32mINFO←[0m Running in Conda
←[32mINFO←[0m Running in a Virtual Environment
←[32mINFO←[0m Encoding: cp1252
←[32mINFO←[0m Installed pip: 24.2
←[32mINFO←[0m Installing packaging
packaging-24.2 | 200 KB | | 0%
←[A packaging-24.2 | 200 KB | ██ | 8%
←[A packaging-24.2 | 200 KB | █████████████████████████████████ | 96%
←[A packaging-24.2 | 200 KB | ███████████████████████████████████ | 100%
←[A packaging-24.2 | 200 KB | ███████████████████████████████████ | 100%
←[32mINFO←[0m Faceswap config written to: C:\Users\\fs3\config\.faceswap
←[32mINFO←[0m Keras config written to: C:\Users\\.keras\keras.json
INFO Installing Required Python Packages. This may take some time...
INFO Installing pywinpty>=2.0.2
ERROR Unable to install package: pywinpty>=2.0.2. Process aborted

[DeliciousCaramels]

Redundant, removed for space.

[BCD16]

Now I have switched from Apple Silicon to a Asus ROG G16 Zephyrus notebook with a RTX 4090 Max-Q (80 watt PL). I installed this beta and forced install newer torch(2.6.0+cu126) and keras(3.3.3).

Now with a batch size of 160 with the original DFLH128 trainer, I can reach up to 255 EGs/sec with only 80 watt GPU power(Mixed precision enabled). Almost 3x faster than old FS2 tensorflow backend. This is insane. With 2000 images input each A and B side, I am almost done in 24 hours. I could give an other 24 hours with "No warp" checked and it comes close to the perfection.

Keep up the good work man! if anyone interested I may upload my fully working conda environment and fs3 app folder. I think it will work for any Nvidia CUDA user under Windows. PM me. Thanks again!

[DeliciousCaramels]

Redundant, space.

[DeliciousCaramels]

Alright so these windows users must be making you lose your mind!

I have made this work now after much trial and error.

Firstly, I follow the basic instructions but before I run the install script I install pytorch into the environment via this command -

conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia

The version of numpy packaged with this is 1.24.3 which causes an error, and so again before running the install script I try installing numpy 1.26.4 as per the requirements, however, this installs a different version of pytorch with cpu in the name, and it won't train. So, I install 1.26.4 via pip.

I check that I have functioning cuda by running the script from this page
https://phphe.com/blog/install-pytorch-manually

Matplotlib needs to be installed manually.

It all installs fine and I can train!

If I train a model started with Conv aware I get this warning

Code: Select all

C:\Users\admin\MiniConda3\envs\fs3\Lib\site-packages\keras\src\backend\torch\nn.py:416: UserWarning: Using padding='same' with even kernel lengths and odd dilation may require a zero-padded copy of the input be created (Triggered internally at C:\cb\pytorch_1000000000000\work\aten\src\ATen\native\Convolution.cpp:1032.)
outputs = tnn.conv2d(

Training iteration speed seems slower in comparison to FS2. In FS2 running my default max settings (phaze-a) I get about 22 eg/s in fs2, in fs3 I'm getting about 9 eg/s.
I get roughly 4000 iterations per hour in fs2 and 2500 in FS3 training using the same settings.

It could be the case that there is some issue with the way I have installed? Googling indicates the warning is due to a mismatch with CUDNN versions but all my cuda packages come from the same source.

Here are my packages and system settings etc., let me know if anything stands out as wrong.

Code: Select all

============ System Information ============
backend:             nvidia
encoding:            cp1252
git_branch:          fs3
git_commits:         8073752 Merge branch 'master' into fs3 | cbaad14 Bugfix: Linux installer - pin git to < 2.45 | 6fe300e pin numpy to < 2.0
gpu_cuda:            No global version found. Check Conda packages for Conda Cuda
gpu_cudnn:           No global version found. Check Conda packages for Conda cuDNN
gpu_devices:         GPU_0: NVIDIA GeForce RTX 3070 Laptop GPU
gpu_devices_active:  GPU_0
gpu_driver:          565.90
gpu_vram:            GPU_0: 8192MB (133MB free)
os_machine:          AMD64
os_platform:         Windows-10-10.0.19045-SP0
os_release:          10
py_conda_version:    conda 25.3.0
py_implementation:   CPython
py_version:          3.11.11
py_virtual_env:      True
sys_cores:           16
sys_processor:       Intel64 Family 6 Model 141 Stepping 1, GenuineIntel
sys_ram:             Total: 32547MB, Available: 13585MB, Used: 18961MB, Free: 13585MB

=============== Pip Packages ===============
absl-py==2.2.0
Brotli @ file:///C:/b/abs_c415aux9ra/croot/brotli-split_1736182803933/work
certifi @ file:///home/conda/feedstock_root/build_artifacts/certifi_1739515848642/work/certifi
charset-normalizer @ file:///croot/charset-normalizer_1721748349566/work
colorama==0.4.6
contourpy @ file:///C:/b/abs_f2u2o_8s9g/croot/contourpy_1732540071787/work
cycler @ file:///tmp/build/80754af9/cycler_1637851556182/work
fastcluster==1.2.6
ffmpy==0.5.0
filelock @ file:///C:/b/abs_f2gie28u58/croot/filelock_1700591233643/work
fonttools @ file:///C:/b/abs_4crkswws2h/croot/fonttools_1737040078745/work
gmpy2 @ file:///C:/b/abs_d8ki0o0h97/croot/gmpy2_1738085498525/work
grpcio==1.71.0
h5py==3.13.0
idna @ file:///C:/b/abs_aad84bnnw5/croot/idna_1714398896795/work
imageio==2.37.0
imageio-ffmpeg==0.6.0
Jinja2 @ file:///C:/b/abs_920kup4e6u/croot/jinja2_1741711580669/work
joblib==1.4.2
keras==3.3.3
kiwisolver @ file:///C:/b/abs_faf90xet7a/croot/kiwisolver_1737040915779/work
Markdown==3.7
markdown-it-py==3.0.0
MarkupSafe @ file:///C:/b/abs_a0ma7ge0jc/croot/markupsafe_1738584052792/work
matplotlib==3.10.0
mdurl==0.1.2
mkl-fft==1.3.1
mkl-random @ file:///C:/ci_311/mkl_random_1676481991689/work
mkl-service==2.4.0
ml_dtypes==0.5.1
mpmath @ file:///C:/b/abs_7833jrbiox/croot/mpmath_1690848321154/work
namex==0.0.8
networkx @ file:///C:/b/abs_b054htfn9t/croot/networkx_1737043671910/work
numexpr==2.10.2
numpy==1.26.4
nvidia-ml-py==12.570.86
opencv-python==4.11.0.86
optree==0.14.1
packaging @ file:///C:/b/abs_3by6s2fa66/croot/packaging_1734472138782/work
pillow @ file:///C:/b/abs_b50vowcrzo/croot/pillow_1738010273782/work
protobuf==6.30.1
psutil==7.0.0
Pygments==2.19.1
pyparsing @ file:///C:/b/abs_40z8gyj9wi/croot/pyparsing_1731445739241/work
PyQt6==6.7.1
PyQt6_sip @ file:///C:/b/abs_28s7k4h_hl/croot/pyqt-split_1740498234166/work/pyqt_sip
PySocks @ file:///C:/ci_311/pysocks_1676425991111/work
python-dateutil @ file:///C:/b/abs_3au_koqnbs/croot/python-dateutil_1716495777160/work
pywin32==310
pywinpty==2.0.15
PyYAML @ file:///C:/b/abs_14xkfs39bx/croot/pyyaml_1728657968772/work
requests @ file:///C:/b/abs_c3508vg8ez/croot/requests_1731000584867/work
rich==13.9.4
scikit-learn==1.6.1
scipy==1.15.2
sip @ file:///C:/b/abs_5cto136kse/croot/sip_1738856220313/work
six @ file:///tmp/build/80754af9/six_1644875935023/work
sympy @ file:///C:/b/abs_b4u17p23yg/croot/sympy_1738108511395/work
tensorboard==2.19.0
tensorboard-data-server==0.7.2
threadpoolctl==3.6.0
torch==2.3.1
torchaudio==2.3.1
torchvision==0.18.1
tornado @ file:///C:/b/abs_7cyu943ybx/croot/tornado_1733960510898/work
tqdm==4.67.1
typing_extensions @ file:///C:/b/abs_0ffjxtihug/croot/typing_extensions_1734714875646/work
unicodedata2 @ file:///C:/b/abs_dfnftvxi4k/croot/unicodedata2_1736543771112/work
urllib3 @ file:///C:/b/abs_7bst06lizn/croot/urllib3_1737133657081/work
Werkzeug==3.1.3
win-inet-pton @ file:///C:/ci_311/win_inet_pton_1676425458225/work

============== Conda Packages ==============
# packages in environment at C:\Users\daniel\MiniConda3\envs\fs3:
#
# Name                    Version                   Build  Channel
absl-py                   2.2.0                    pypi_0    pypi
blas                      1.0                         mkl  

brotli-python             1.0.9           py311h5da7b33_9  

bzip2                     1.0.8                h2bbff1b_6  

ca-certificates           2025.2.25            haa95532_0  

certifi                   2025.1.31          pyhd8ed1ab_0    conda-forge
charset-normalizer        3.3.2              pyhd3eb1b0_0  

colorama                  0.4.6                    pypi_0    pypi
contourpy                 1.3.1           py311h214f63a_0  

cuda-cccl                 12.8.90                       0    nvidia
cuda-cccl_win-64          12.8.90                       0    nvidia
cuda-cudart               12.1.105                      0    nvidia
cuda-cudart-dev           12.1.105                      0    nvidia
cuda-cupti                12.1.105                      0    nvidia
cuda-libraries            12.1.0                        0    nvidia
cuda-libraries-dev        12.1.0                        0    nvidia
cuda-nvrtc                12.1.105                      0    nvidia
cuda-nvrtc-dev            12.1.105                      0    nvidia
cuda-nvtx                 12.1.105                      0    nvidia
cuda-opencl               12.8.90                       0    nvidia
cuda-opencl-dev           12.8.90                       0    nvidia
cuda-profiler-api         12.8.90                       0    nvidia
cuda-runtime              12.1.0                        0    nvidia
cuda-version              12.8                          3    nvidia
cycler                    0.11.0             pyhd3eb1b0_0  

fastcluster               1.2.6                    pypi_0    pypi
ffmpy                     0.5.0                    pypi_0    pypi
filelock                  3.13.1          py311haa95532_0  

fonttools                 4.55.3          py311h827c3e9_0  

freetype                  2.12.1               ha860e81_0  

git                       2.45.2               haa95532_1  

gmp                       6.3.0                h537511b_0  

gmpy2                     2.2.1           py311h827c3e9_0  

grpcio                    1.71.0                   pypi_0    pypi
h5py                      3.13.0                   pypi_0    pypi
icu                       73.1                 h6c2663c_0  

idna                      3.7             py311haa95532_0  

imageio                   2.37.0                   pypi_0    pypi
imageio-ffmpeg            0.6.0                    pypi_0    pypi
intel-openmp              2021.4.0          haa95532_3556  

jinja2                    3.1.6           py311haa95532_0  

joblib                    1.4.2                    pypi_0    pypi
jpeg                      9e                   h827c3e9_3  

keras                     3.3.3                    pypi_0    pypi
khronos-opencl-icd-loader 2024.05.08           h8cc25b3_0  

kiwisolver                1.4.8           py311h5da7b33_0  

krb5                      1.20.1               h5b6d351_0  

lcms2                     2.16                 hb4a4139_0  

lerc                      4.0.0                h5da7b33_0  

libcublas                 12.1.0.26                     0    nvidia
libcublas-dev             12.1.0.26                     0    nvidia
libcufft                  11.0.2.4                      0    nvidia
libcufft-dev              11.0.2.4                      0    nvidia
libcurand                 10.3.9.90                     0    nvidia
libcurand-dev             10.3.9.90                     0    nvidia
libcusolver               11.4.4.55                     0    nvidia
libcusolver-dev           11.4.4.55                     0    nvidia
libcusparse               12.0.2.55                     0    nvidia
libcusparse-dev           12.0.2.55                     0    nvidia
libdeflate                1.22                 h5bf469e_0  

libffi                    3.4.4                hd77b12b_1  

libjpeg-turbo             2.0.0                h196d8e1_0  

libnpp                    12.0.2.50                     0    nvidia
libnpp-dev                12.0.2.50                     0    nvidia
libnvjitlink              12.1.105                      0    nvidia
libnvjitlink-dev          12.1.105                      0    nvidia
libnvjpeg                 12.1.1.14                     0    nvidia
libnvjpeg-dev             12.1.1.14                     0    nvidia
libpng                    1.6.39               h8cc25b3_0  

libpq                     17.4                 h70ee33d_0  

libtiff                   4.5.1                h44ae7cf_1  

libuv                     1.48.0               h827c3e9_0  

libwebp-base              1.3.2                h3d04722_1  

libzlib                   1.2.13               hcfcfb64_4    conda-forge
libzlib-wapi              1.2.13               hcfcfb64_4    conda-forge
lz4-c                     1.9.4                h2bbff1b_1  

markdown                  3.7                      pypi_0    pypi
markdown-it-py            3.0.0                    pypi_0    pypi
markupsafe                3.0.2           py311h827c3e9_0  

matplotlib                3.10.0          py311haa95532_0  

matplotlib-base           3.10.0          py311he19b0ae_0  

mdurl                     0.1.2                    pypi_0    pypi
mkl                       2021.4.0           haa95532_640  

mkl-service               2.4.0           py311h2bbff1b_0  

mkl_fft                   1.3.1           py311h743a336_0  

mkl_random                1.2.2           py311heda8569_0  

ml-dtypes                 0.5.1                    pypi_0    pypi
mpc                       1.3.1                h827c3e9_0  

mpfr                      4.2.1                h56c3642_0  

mpmath                    1.3.0           py311haa95532_0  

namex                     0.0.8                    pypi_0    pypi
networkx                  3.4.2           py311haa95532_0  

numexpr                   2.10.2                   pypi_0    pypi
numpy                     1.26.4                   pypi_0    pypi
nvidia-ml-py              12.570.86                pypi_0    pypi
opencv-python             4.11.0.86                pypi_0    pypi
openjpeg                  2.5.2                hae555c5_0  

openssl                   3.1.0                hcfcfb64_3    conda-forge
optree                    0.14.1                   pypi_0    pypi
packaging                 24.2            py311haa95532_0  

pillow                    11.1.0          py311h096bfcc_0  

pip                       25.0            py311haa95532_0  

protobuf                  6.30.1                   pypi_0    pypi
psutil                    7.0.0                    pypi_0    pypi
pygments                  2.19.1                   pypi_0    pypi
pyparsing                 3.2.0           py311haa95532_0  

pyqt                      6.7.1           py311h5da7b33_0  

pyqt6-sip                 13.9.1          py311h827c3e9_0  

pysocks                   1.7.1           py311haa95532_0  

python                    3.11.11              h4607a30_0  

python-dateutil           2.9.0post0      py311haa95532_2  

pytorch                   2.3.1           py3.11_cuda12.1_cudnn8_0    pytorch
pytorch-cuda              12.1                 hde6ce7c_6    pytorch
pytorch-mutex             1.0                        cuda    pytorch
pywin32                   310                      pypi_0    pypi
pyyaml                    6.0.2           py311h827c3e9_0  

qtbase                    6.7.2                h0804d20_1  

qtdeclarative             6.7.2                h5da7b33_0  

qtsvg                     6.7.2                hf2fb9eb_0  

qttools                   6.7.2                h0de5f00_0  

qtwebchannel              6.7.2                h5da7b33_0  

qtwebsockets              6.7.2                h5da7b33_0  

requests                  2.32.3          py311haa95532_1  

rich                      13.9.4                   pypi_0    pypi
scikit-learn              1.6.1                    pypi_0    pypi
scipy                     1.15.2                   pypi_0    pypi
setuptools                75.8.0          py311haa95532_0  

sip                       6.10.0          py311h5da7b33_0  

six                       1.16.0             pyhd3eb1b0_1  

sqlite                    3.45.3               h2bbff1b_0  

sympy                     1.13.3          py311haa95532_1  

tensorboard               2.19.0                   pypi_0    pypi
tensorboard-data-server   0.7.2                    pypi_0    pypi
threadpoolctl             3.6.0                    pypi_0    pypi
tk                        8.6.14               h0416ee5_0  

torchaudio                2.3.1                    pypi_0    pypi
torchvision               0.18.1                   pypi_0    pypi
tornado                   6.4.2           py311h827c3e9_0  

tqdm                      4.67.1                   pypi_0    pypi
typing_extensions         4.12.2          py311haa95532_0  

tzdata                    2025a                h04d1e81_0  

ucrt                      10.0.22621.0         h57928b3_1    conda-forge
unicodedata2              15.1.0          py311h827c3e9_1  

urllib3                   2.3.0           py311haa95532_0  

vc                        14.42                haa95532_4  

vs2015_runtime            14.42.34433          he0abc0d_4  

werkzeug                  3.1.3                    pypi_0    pypi
wheel                     0.45.1          py311haa95532_0  

win_inet_pton             1.1.0           py311haa95532_0  

xz                        5.6.4                h4754444_1  

yaml                      0.2.5                he774522_0  

zlib                      1.2.13               hcfcfb64_4    conda-forge
zlib-wapi                 1.2.13               hcfcfb64_4    conda-forge
zstd                      1.5.6                h8880b57_0  

On the plus side, I have a model running in effect the dny256 presets with efficientnetv2/b3 encoder, and at 150k iterations running mixed precision I got a NAN in fs2 (4 learning rate, adabelief, -13 epsilon), and even rolling back to 130k iterations I couldn't run training speed over 1.5 with MP disabled, it would NAN after about 100 iterations so I abandoned the model.

Loaded the 150k iteration version into FS3, MP, 4e5 learning rate, so far no issues after 1k iterations and so there seems to be some benefit to the new back end.
**Edit. As you would expect, despite there being no NANs, the model is ruined and after 5000 iterations the images look cursed, purple and green. Interesting though that no NANs occurred, could it be the case the NAN protection is not working in FS3?

**edit two : started a new model using efficientnetb3 with the RMS-Prop optimiser (a paper states its the most accurate model with b3? Give it a go!)
Epsilon at -4, learning rate 4e5 and MP enabled, got a NAN in less than 10k iterations. Wild!

Another thing, it appears there is some error with DFL-SAE model compatibility; created a new model at 256 res and all other settings standard, batch size 4. Training speed is extremely slow, it runs about 1 iteration every 3 seconds, and exiting training causes a crash. I can investigate/ provide more info on this but I would suspect it relates to an incorrect package?

Edit three : Thinking about it, there are a lot more options for the optimizers. I get much better training rates (about 18, and I can run a batch size of 8) running RMSProp which I believe is because Adabelief is relatively vram heavy. Is it plausible that the slower training rate with my "default" settings is due to the additional optimiser options?

Edit four: RMSprop needs some more research, nothing but NANS and the rate of actual change is woeful, 10k iterations was my best go today and it looked nothing like the results I have been getting with adab/ Efficient, not even to the point of the "hollowed out" faces.
Have tried everything from EEs up to -8, learning rates down to 1; turned on gradient clipping. So I instead tried AdamW, so far after a mere 1k iterations I had the hollowed out/ pumpkin head looking faces. After the 1250 iteration model update I have the blown out/ bright red and yellow faces. Is this model collapsing already?

A lot of interesting differences going on here in general. I hope all users get a chance to play with this and hopefully someone with a little more computing power than I have finds some cool stuff

[DeliciousCaramels]

Just to help people not have to dig through my post here are further installation instructions that worked for me on Windows 10 / nvidia backend.

As you follow Torzdfs instructions; when you get past the part that you activate your Conda environment (ie, you are in your "fs3" environment) add the following commands before you run the install script.

conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia

conda install matplotlib

pip install numpy==1.26.4

conda install pywinpty

Then run install script.

I just tested this on a system that doesn't have any other faceswaps installed, ie, a brand new Conda install etc. and it works.

Good luck all and have fun!

[DeliciousCaramels]

It seems there may be some issues. Not certain at this point as obviously this is a very customisable piece of software. However, I think they relate to vram allocation.

1) Installed FS3 on my "extraction" box (it has an 11gb 1080ti). FS2 will run parallel extraction with 7 batch size s3fd and 12 fan align; this assigns about 9.1gb of the vram to do this.
FS3 I can only run 3/10 and it will only assign about 5.9gb of vram. I also tested both using identical settings (ie, I tried extraction in FS2 with 3/10 batch sizes) and the actual extraction itself is about 10-15% slower in FS3.

2) Conversely, FS3 training will assign more Vram. I have an 8gb 3070 in my "training" laptop and FS3 will assign just under 7.3gb of Vram at max, compared to 7.1gb in regular. This does allow me to run slightly higher batch sizes.

3) Every model I have made leads to crashing/ Nans, and it happens quickly. So far the most long lasting model was 35k iterations, however the model itself started gaining loss after 15k iterations and at 27k the B face loss jumped immediately from 0.3 to 2.0! Just playing with settings, loss functions etc. Current model I am training the loss has again started to climb after 10k iterations.

Of course in all cases I am running mixed precision. Perhaps tomorrow I will start a model without it and see what happens.

*edit.
Well it didn't take long. To avoid any potential issues with phaze-a settings I started a new model, DFL-SAE at 160 with the DF arch. Without MP I can get a batch size of 2, which is lower than what I could get on my old laptop with a 6gb 1070 (could run at 3), and I got a NAN warning in 768 iterations.

*Edit 2.
Ok, well, problem solved. I'm pretty sure that LPIPS is an issue here, which makes sense given that this environment has newer cudatoolkits relative to my FS2 install. Will start a new phaze-a with just MS-SSIM and MAE and see how it goes.

*Edit 3.
Training VRAM allocation is significantly better, I can run my regular model at a batch size of 10 where in FS2 I can only do 8. It will actually allocate up to 7.7gb of my 8gb Vram, very impressed. So far dumping LPIPs I have 60k iterations in a new model today and no problems at all, getting a faster training rate than ever before, 28 eg/s on Phaze a/ effnetv2/b3.

Extraction speed issue and LPIPS redundancy aside, I highly recommend people install this. Excellent work !!!

[DeliciousCaramels]

Thought I would try a quick convert to see how the new model is looking and I came up with an error - I tried to run preview mode and the output was slightly more verbose.

Code: Select all

Traceback (most recent call last):
  File "D:\fs3\lib\convert.py", line 203, in process
    image = self._patch_image(item)
            ^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\fs3\lib\convert.py", line 267, in _patch_image
    new_image, background = self._get_new_image(predicted, frame_size)
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\fs3\lib\convert.py", line 361, in _get_new_image
    predicted_mask = new_face[:, :, -1] if new_face.shape[2] == 4 else None
                                           ~~~~~~~~~~~~~~^^^IndexError: tuple index out of range

I'm not running predicted mask at all, I don't run learn mask cause I found it incompatible with "conv aware init".

[DeliciousCaramels]

Been playing around again. I found on another model that the reason I was getting NaNs using LPIPS on my model was some of the faces were not well aligned. I understand that alignment doesn't have anything to do with training but removing about 500 images that had messed up alignments (even though faces appeared masked OK the obscuring or angling of the face meant the aligner was just a blob of points) allowed me to continue training it, so I thought I'd give FS3 another go.

Currently it seems running the dataset that was causing immediate NaN's previously absent the poorly aligned images allows training, no problems at all! So there must have been some images that for whatever reason when fed into the encoder just make it nope.

I've started new models to test back to back on FS2 and FS3 and same settings I can run FS3 at 7 i/s vs 4. I'm going to let FS3 run this model as long as possible now to see if I get NaNs as it progresses.

[BCD16]

Hello everyone. I am adding my own Phaze-A 192px preset below. This preset was written by me on the basis of DeepFaceLab's original SAEHD LIAE-UDT preset. Save it as a .json file and load in Phaze-A. Use %100 face covering while training.

Use Bisenet-Fp masker for both A and B faces while extraction. Make sure you use same masker while training and "Warp To Landmarks" option enabled.

Make sure Mask type in Settings-Train-Global-Loss page selected as "Bisenet-Fp Face"

I am getting extremely good results. My GPU is a RTX 4090, and I can train 180 EGs/sec with a batch size 64 (mixed precision enabled). It is extremely good for a 192px preset.
Some packages I force upgraded and works well with FS3 beta branch. (For NVIDIA RTX 3000 and 4000 series these should work)

keras 3.3.3 pypi_0 pypi
numpy 1.26.4 pypi_0 pypi
nvidia-ml-py 12.570.86 pypi_0 pypi
torch 2.8.0.dev20250327+cu128 pypi_0 pypi
torchaudio 2.6.0.dev20250401+cu128 pypi_0 pypi
torchvision 0.22.0.dev20250401+cu128 pypi_0 pypi

Code: Select all

{
  "output_size": 192,
  "shared_fc": "half",
  "enable_gblock": false,
  "split_fc": true,
  "split_gblock": false,
  "split_decoders": false,
  "enc_architecture": "fs_original",
  "enc_scaling": 13,
  "enc_load_weights": false,
  "bottleneck_type": "dense",
  "bottleneck_norm": "none",
  "bottleneck_size": 256,
  "bottleneck_in_encoder": false,
  "fc_depth": 1,
  "fc_min_filters": 512,
  "fc_max_filters": 512,
  "fc_dimensions": 8,
  "fc_filter_slope": -0.5,
  "fc_dropout": 0.0,
  "fc_upsampler": "subpixel",
  "fc_upsamples": 1,
  "fc_upsample_filters": 512,
  "fc_gblock_depth": 3,
  "fc_gblock_min_nodes": 512,
  "fc_gblock_max_nodes": 512,
  "fc_gblock_filter_slope": -0.5,
  "fc_gblock_dropout": 0.0,
  "dec_upscale_method": "subpixel",
  "dec_upscales_in_fc": 0,
  "dec_norm": "none",
  "dec_min_filters": 64,
  "dec_max_filters": 512,
  "dec_slope_mode": "full",
  "dec_filter_slope": -0.33,
  "dec_res_blocks": 1,
  "dec_output_kernel": 1,
  "dec_gaussian": false,
  "dec_skip_last_residual": false,
  "freeze_layers": "encoder",
  "load_layers": "encoder",
  "fs_original_depth": 4,
  "fs_original_min_filters": 64,
  "fs_original_max_filters": 512,
  "fs_original_use_alt": false,
  "mobilenet_width": 1.0,
  "mobilenet_depth": 1,
  "mobilenet_dropout": 0.001,
  "__filetype": "faceswap_preset",
  "__section": "train|model|phaze_a"
}

[DeliciousCaramels]

Decided to play around a bit with extract to see if I could generate some useful testing information.

I note something interesting. In FS2 I can extract in parallel, using S3FD/ FAN, with batch sizes 5/14 respectively, using refeed setting of 4. I apply masks after extraction as I like to clean up the alignments first.

In FS3, it will only extract in parallel at 1/5. I cannot increase either the detector or aligner even a single BS point higher.
In single process, I can bump FS2 to 5/16, which does give me more stability for longer videos.

However in FS3, I can run detect at BS6, but the aligner will run out of memory/ crash if I set it above 5.
Interestingly, when I set single process in FS3, the process starts much faster than FS2; it begins detection within 10 seconds, whereas FS2 takes a good 30 or 40 seconds generally.

It seems to me from testing that the FAN aligner isn't working well with the new backend and I think it is something to do with VRAM allocation.

I did some A/B testing using parallel processing, and FS3's maximum setting of 1/5, and just let both extract the first 400 frames of the same video - all settings equal.

FS2 extraction took 1:50, during which time it consumed 6.7gb of VRAM

FS3 extraction took 2:59, during which time it consumed 4.6gb of VRAM.

Further, I did the same test running single process with the detector at a batch size of 5.

FS2 detected 400 frames in 23 seconds, again consuming 6.7gb of VRAM.

FS3 took 18 seconds and consumed only 6gb of VRAM. So detector works better!

I hope this is helpful.