Maximum resolution is increased to 640.
‘hd’ archi is removed. ‘hd’ was experimental archi created to remove subpixel shake, but ‘lr_dropout’ and ‘disable random warping’ do that better.
‘uhd’ is renamed to ‘-u’
dfuhd and liaeuhd will be automatically renamed to df-u and liae-u in existing models.
Added new experimental archi (key -d) which doubles the resolution using the same computation cost.
It is mean same configs will be x2 faster, or for example you can set 448 resolution and it will train as 224.
Strongly recommended not to train from scratch and use pretrained models.
New archi naming:
'df' keeps more identity-preserved face.
'liae' can fix overly different face shapes.
'-u' increased likeness of the face.
'-d' (experimental) doubling the resolution using the same computation cost
Examples: df, liae, df-d, df-ud, liae-ud, ...
Improved GAN training (GAN_power option). It was used for dst model, but actually we don’t need it for dst.
Instead, a second src GAN model with x2 smaller patch size was added, so the overall quality for hi-res models should be higher.
Added option ‘Uniform yaw distribution of samples (y/n)’:
Helps to fix blurry side faces due to small amount of them in the faceset.
Quick96:
Now based on df-ud archi and 20% faster.
XSeg trainer:
Improved sample generator.
Now it randomly adds the background from other samples.
Result is reduced chance of random mask noise on the area outside the face.
Now you can specify ‘batch_size’ in range 2-16.
Reduced size of samples with applied XSeg mask. Thus size of packed samples with applied xseg mask is also reduced.
Fixed "Write preview history". Now it writes all subpreviews in separated folders
https://i.imgur.com/IszifCJ.jpg
also the last preview saved as _last.jpg before the first file
https://i.imgur.com/Ls1AOK4.jpg
thus you can easily check the changes with the first file in photo viewer
with XSeg model you can train your own mask segmentator of dst(and src) faces
that will be used in merger for whole_face.
Instead of using a pretrained model (which does not exist),
you control which part of faces should be masked.
Workflow is not easy, but at the moment it is the best solution
for obtaining the best quality of whole_face's deepfakes using minimum effort
without rotoscoping in AfterEffects.
new scripts:
XSeg) data_dst edit.bat
XSeg) data_dst merge.bat
XSeg) data_dst split.bat
XSeg) data_src edit.bat
XSeg) data_src merge.bat
XSeg) data_src split.bat
XSeg) train.bat
Usage:
unpack dst faceset if packed
run XSeg) data_dst split.bat
this scripts extracts (previously saved) .json data from jpg faces to use in label tool.
run XSeg) data_dst edit.bat
new tool 'labelme' is used
use polygon (CTRL-N) to mask the face
name polygon "1" (one symbol) as include polygon
name polygon "0" (one symbol) as exclude polygon
'exclude polygons' will be applied after all 'include polygons'
Hot keys:
ctrl-N create polygon
ctrl-J edit polygon
A/D navigate between frames
ctrl + mousewheel image zoom
mousewheel vertical scroll
alt+mousewheel horizontal scroll
repeat for 10/50/100 faces,
you don't need to mask every frame of dst,
only frames where the face is different significantly,
for example:
closed eyes
changed head direction
changed light
the more various faces you mask, the more quality you will get
Start masking from the upper left area and follow the clockwise direction.
Keep the same logic of masking for all frames, for example:
the same approximated jaw line of the side faces, where the jaw is not visible
the same hair line
Mask the obstructions using polygon with name "0".
run XSeg) data_dst merge.bat
this script merges .json data of polygons into jpg faces,
therefore faceset can be sorted or packed as usual.
run XSeg) train.bat
train the model
Check the faces of 'XSeg dst faces' preview.
if some faces have wrong or glitchy mask, then repeat steps:
split
run edit
find these glitchy faces and mask them
merge
train further or restart training from scratch
Restart training of XSeg model is only possible by deleting all 'model\XSeg_*' files.
If you want to get the mask of the predicted face in merger,
you should repeat the same steps for src faceset.
New mask modes available in merger for whole_face:
XSeg-prd - XSeg mask of predicted face -> faces from src faceset should be labeled
XSeg-dst - XSeg mask of dst face -> faces from dst faceset should be labeled
XSeg-prd*XSeg-dst - the smallest area of both
if workspace\model folder contains trained XSeg model, then merger will use it,
otherwise you will get transparent mask by using XSeg-* modes.
Some screenshots:
label tool: https://i.imgur.com/aY6QGw1.jpg
trainer : https://i.imgur.com/NM1Kn3s.jpg
merger : https://i.imgur.com/glUzFQ8.jpg
example of the fake using 13 segmented dst faces
: https://i.imgur.com/wmvyizU.gifv
SAEHD:
added new option
GAN power 0.0 .. 10.0
Train the network in Generative Adversarial manner.
Forces the neural network to learn small details of the face.
You can enable/disable this option at any time,
but better to enable it when the network is trained enough.
Typical value is 1.0
GAN power with pretrain mode will not work.
Example of enabling GAN on 81k iters +5k iters
https://i.imgur.com/OdXHLhU.jpghttps://i.imgur.com/CYAJmJx.jpg
dfhd: default Decoder dimensions are now 48
the preview for 256 res is now correctly displayed
fixed model naming/renaming/removing
Improvements for those involved in post-processing in AfterEffects:
Codec is reverted back to x264 in order to properly use in AfterEffects and video players.
Merger now always outputs the mask to workspace\data_dst\merged_mask
removed raw modes except raw-rgb
raw-rgb mode now outputs selected face mask_mode (before square mask)
'export alpha mask' button is replaced by 'show alpha mask'.
You can view the alpha mask without recompute the frames.
8) 'merged *.bat' now also output 'result_mask.' video file.
8) 'merged lossless' now uses x264 lossless codec (before PNG codec)
result_mask video file is always lossless.
Thus you can use result_mask video file as mask layer in the AfterEffects.
Removed the wait at first launch for most graphics cards.
Increased speed of training by 10-20%, but you have to retrain all models from scratch.
SAEHD:
added option 'use float16'
Experimental option. Reduces the model size by half.
Increases the speed of training.
Decreases the accuracy of the model.
The model may collapse or not train.
Model may not learn the mask in large resolutions.
true_face_training option is replaced by
"True face power". 0.0000 .. 1.0
Experimental option. Discriminates the result face to be more like the src face. Higher value - stronger discrimination.
Comparison - https://i.imgur.com/czScS9q.png
If you want, you can manually remove unnecessary angles from src faceset after sort by yaw.
Optimized sample generators (CPU workers). Now they consume less amount of RAM and work faster.
added
4.2.other) data_src/dst util faceset pack.bat
Packs /aligned/ samples into one /aligned/samples.pak file.
After that, all faces will be deleted.
4.2.other) data_src/dst util faceset unpack.bat
unpacks faces from /aligned/samples.pak to /aligned/ dir.
After that, samples.pak will be deleted.
Packed faceset load and work faster.
Random warp is required to generalize facial expressions of both faces. When the face is trained enough, you can disable it to get extra sharpness for less amount of iterations.
added SAEHD model ( High Definition Styled AutoEncoder )
This is a new heavyweight model for high-end cards to achieve maximum possible deepfake quality in 2020.
Differences from SAE:
+ new encoder produces more stable face and less scale jitter
before: https://i.imgur.com/4jUcol8.gifv
after: https://i.imgur.com/lyiax49.gifv - scale of the face is less changed within frame size
+ new decoder produces subpixel clear result
+ pixel loss and dssim loss are merged together to achieve both training speed and pixel trueness
+ by default networks will be initialized with CA weights, but only after first successful iteration
therefore you can test network size and batch size before weights initialization process
+ new neural network optimizer consumes less VRAM than before
+ added option <Enable 'true face' training>
The result face will be more like src and will get extra sharpness.
example: https://i.imgur.com/ME3A7dI.gifv
Enable it for last 15-30k iterations before conversion.
+ encoder and decoder dims are merged to one parameter encoder/decoder dims
+ added mid-full face, which covers 30% more area than half face.
fixed model sizes from previous update.
avoided bug in ML framework(keras) that forces to train the model on random noise.
Converter: added blur on the same keys as sharpness
Added new model 'TrueFace'. This is a GAN model ported from https://github.com/NVlabs/FUNIT
Model produces near zero morphing and high detail face.
Model has higher failure rate than other models.
Keep src and dst faceset in same lighting conditions.
Session is now saved to the model folder.
blur and erode ranges are increased to -400+400
hist-match-bw is now replaced with seamless2 mode.
Added 'ebs' color transfer mode (works only on Windows).
FANSEG model (used in FAN-x mask modes) is retrained with new model configuration
and now produces better precision and less jitter
With interactive converter you can change any parameter of any frame and see the result in real time.
Converter: added motion_blur_power param.
Motion blur is applied by precomputed motion vectors.
So the moving face will look more realistic.
RecycleGAN model is removed.
Added experimental AVATAR model. Minimum required VRAM is 6GB (NVIDIA), 12GB (AMD)
Usage:
1) place data_src.mp4 10-20min square resolution video of news reporter sitting at the table with static background,
other faces should not appear in frames.
2) process "extract images from video data_src.bat" with FULL fps
3) place data_dst.mp4 video of face who will control the src face
4) process "extract images from video data_dst FULL FPS.bat"
5) process "data_src mark faces S3FD best GPU.bat"
6) process "data_dst extract unaligned faces S3FD best GPU.bat"
7) train AVATAR.bat stage 1, tune batch size to maximum for your card (32 for 6GB), train to 50k+ iters.
8) train AVATAR.bat stage 2, tune batch size to maximum for your card (4 for 6GB), train to decent sharpness.
9) convert AVATAR.bat
10) converted to mp4.bat
updated versions of modules
An issue affecting at least 2070 and 2080 cards (possibly other RTX cards too) requires auto growth to be enabled for TensorFlow to work.
I don't know enough about the impact of this change to know whether this ought to be made optional or not, but for RTX owners, this simple change fixes TensorFlow errors when generating models.
Enable autobackup? (y/n ?:help skip:%s) :
Autobackup model files with preview every hour for last 15 hours. Latest backup located in model/<>_autobackups/01
SAE: added option only for CUDA builds:
Enable gradient clipping? (y/n, ?:help skip:%s) :
Gradient clipping reduces chance of model collapse, sacrificing speed of training.
Pretrain the model with large amount of various faces. This technique may help to train the fake with overly different face shapes and light conditions of src/dst data. Face will be look more like a morphed. To reduce the morph effect, some model files will be initialized but not be updated after pretrain: LIAE: inter_AB.h5 DF: both decoders.h5. The longer you pretrain the model the more morphed face will look. After that, save and run the training again.
Pixel loss may help to enhance fine details and stabilize face color. Use it only if quality does not improve over time.
SAE:
previous SAE model will not work with this update.
Greatly decreased chance of model collapse.
Increased model accuracy.
Residual blocks now default and this option has been removed.
Improved 'learn mask'.
Added masked preview (switch by space key)
Converter:
fixed rct/lct in seamless mode
added mask mode (6) learned*FAN-prd*FAN-dst
added mask editor, its created for refining dataset for FANSeg model, and not for production, but you can spend your time and test it in regular fakes with face obstructions
