mirror of
https://github.com/iperov/DeepFaceLab.git
synced 2025-07-12 16:13:51 -07:00
164 lines
6.2 KiB
Markdown
164 lines
6.2 KiB
Markdown
## **DeepFaceLab** is a tool that utilizes deep learning to recognize and swap faces in pictures and videos.
|
|
|
|
Based on original FaceSwap repo. **Facesets** of FaceSwap or FakeApp are **not compatible** with this repo. You should to run extract again.
|
|
|
|
### **Features**:
|
|
|
|
- new models
|
|
|
|
- new architecture, easy to experiment with models
|
|
|
|
- works on 2GB old cards , such as GT730. Example of fake trained on 2GB gtx850m notebook in 18 hours https://www.youtube.com/watch?v=bprVuRxBA34
|
|
|
|
- face data embedded to png files
|
|
|
|
- automatic GPU manager, chooses best gpu(s) and supports --multi-gpu (only for identical cards)
|
|
|
|
- new preview window
|
|
|
|
- extractor in parallel
|
|
|
|
- converter in parallel
|
|
|
|
- added **--debug** option for all stages
|
|
|
|
- added **MTCNN extractor** which produce less jittered aligned face than DLIBCNN, but can produce more false faces. Comparison dlib (at left) vs mtcnn on hard case:
|
|

|
|
MTCNN produces less jitter.
|
|
|
|
- added **Manual extractor**. You can fix missed faces manually or do full manual extract:
|
|

|
|

|
|
|
|
- standalone zero dependencies ready to work prebuilt binary for all windows versions, see below
|
|
|
|
### **Model types**:
|
|
|
|
- **H64 (2GB+)** - half face with 64 resolution. It is as original FakeApp or FaceSwap, but with new TensorFlow 1.8 DSSIM Loss func and separated mask decoder + better ConverterMasked. for 2GB and 3GB VRAM model works in reduced mode.
|
|
|
|
H64 Robert Downey Jr.:
|
|
|
|

|
|
|
|

|
|
|
|
- **H128 (3GB+)** - as H64, but in 128 resolution. Better face details. for 3GB and 4GB VRAM model works in reduced mode.
|
|
|
|
H128 Cage:
|
|
|
|

|
|
|
|
H128 asian face on blurry target:
|
|
|
|

|
|
|
|

|
|
|
|
- **DF (5GB+)** - @dfaker model. As H128, but fullface model. Strongly recommended not to mix various light conditions in src faces.
|
|
|
|

|
|
|
|
- **LIAEF128 (5GB+)** - new model. Result of combining DF, IAE, + experiments. Model tries to morph src face to dst, while keeping facial features of src face, but less agressive morphing. Model has problems with closed eyes recognizing.
|
|
|
|
LIAEF128 Cage:
|
|
|
|

|
|
|
|

|
|
|
|
LIAEF128 Cage video:
|
|
|
|
[](https://www.youtube.com/watch?v=mRsexePEVco)
|
|
|
|
- **LIAEF128YAW (5GB+)** - currently testing. Useful when your src faceset has too many side faces vs dst faceset. It feeds NN by sorted samples by yaw.
|
|
|
|
- **MIAEF128 (5GB+)** - as LIAEF128, but also it tries to match brightness/color features.
|
|
|
|
MIAEF128 model diagramm:
|
|
|
|

|
|
|
|
MIAEF128 Ford success case:
|
|
|
|

|
|
|
|
MIAEF128 Cage fail case:
|
|
|
|

|
|
|
|
- **AVATAR (4GB+)** - non GAN, 256x256 face controlling model.
|
|
|
|

|
|
|
|
Video:
|
|
|
|
[](https://www.youtube.com/watch?v=3M0E4QnWMqA)
|
|
|
|
Usage:
|
|
|
|
src - controllable face (Cage)
|
|
|
|
dst - controller face (your face)
|
|
|
|
converter --input-dir must contains *extracted dst faces* in sequence to be converted, its mean you can train on 1500 dst faces, but use only 100 for convert.
|
|
|
|
- Video comparison of different Cage facesets.
|
|
Vertical: 1 - mix of various Cage face shape and light conditions. 2,3,4 - without mix.
|
|
Horizontal: 1 - DF, 2 - LIAEF128.
|
|
|
|
[](https://youtu.be/C1nFgrmtm_o)
|
|
|
|
Conclusion: **better not to mix and use only same shape faces with same light**
|
|
|
|
### **Sort tool**:
|
|
|
|
`hist` groups images by similar content
|
|
|
|
`hist-dissim` places most similar to each other images to end.
|
|
|
|
`hist-blur` sort by blur in groups of similar content
|
|
|
|
`brightness`
|
|
|
|
`hue`
|
|
|
|
`face` and `face-dissim` currently useless
|
|
|
|
Best practice for gather src faceset:
|
|
|
|
1) delete first unsorted aligned groups of images what you can to delete. Dont touch target face mixed with others.
|
|
2) `blur` -> delete ~half of them
|
|
3) `hist` -> delete groups of similar and leave only target face
|
|
4) `hist-blur` -> delete blurred at end of groups of similar
|
|
5) `hist-dissim` -> leave only first **1000-1500 faces**, because number of src faces can affect result. For YAW feeder model skip this step.
|
|
6) `face-yaw` -> just for finalize faceset
|
|
|
|
Best practice for dst faces:
|
|
|
|
1) delete first unsorted aligned groups of images what you can to delete. Dont touch target face mixed with others.
|
|
2) `hist` -> delete groups of similar and leave only target face
|
|
|
|
### **Facesets**:
|
|
|
|
- Nicolas Cage 4 facesets (1 mix + 3 different)
|
|
|
|
download from here: https://mega.nz/#F!y1ERHDaL!PPwg01PQZk0FhWLVo5_MaQ
|
|
|
|
### **Build info**
|
|
|
|
dlib==19.10.0 from pip compiled without CUDA. Therefore you have to compile DLIB manually.
|
|
|
|
Command line example for windows: `python setup.py install -G "Visual Studio 14 2015" --yes DLIB_USE_CUDA`
|
|
|
|
### **Prebuilt python folder with DeepFaceLab**:
|
|
|
|
Windows 7,8,8.1,10 zero dependency (just install/update your GeForce Drivers) prebuilt Python 3.6.5 embeddable folder with DeepFaceLab can be downloaded from torrent https://rutracker.org/forum/viewtopic.php?p=75318742 (magnet link inside).
|
|
|
|
### **Windows 10 memory problem:
|
|
|
|
Windows 10 consumes % of VRAM even if card unused for video output.
|
|
|
|
### **Pull requesting**:
|
|
|
|
I understand some people want to help. But result of mass people contribution we can see in deepfakes\faceswap.
|
|
High chance I will decline PR. Therefore before PR better ask me what you want to change or add to save your time.
|