Hi guys, I used DFL last year and did not care anymore because results were not as expected. I am impressed what have been done so far and the DF topic catched my mind again when sawing the home stallone fake on youtube.
I have to say: I want to make a deepfake with model src and pornstar dst and body simmilarity is most important so im limited in choosing the right face structre. but my dst pornstar is pretty close, the upper half of the face is similar to dst, the lower half is bigger, bigger chin, bigger mouth, to mention that
now I went by the old guide in the DFL 1.0 outdated thread 'My Personal DeepFake Workflow Using DeepFaceLab' from dpfks. I already went hard on sorting just the best src images, min res. 720p and only sharp images without shadows etc I think I did a pretty good job on this one.
the dst was harder for me because I have angles on the pornstar where camera is looking on the head pointing on the forehead or very from bottm pointing at the (double)chin. landmarks on these were faulty so I re extracetd manually but even here some of those are simply not correct landmarked no matter how I scroll or move the mouse. I tried the nearest and saved it.
I need to mention: I have very much angles of the src model from interview clips and such, but getting those very up and bottom angles is very hard to find so actually I dont have those very unusual angles in src. first question here is: If I find some how can I add them to my src and continue training without restart the whole model?
Now for settings atm I am using:
at start I had batch size 8, unfortunately no clipgrad enabled, gan power 0, trueface 0, face style 10 bg style 10. changed at about 80k its like mentioned in the old HowTo. Now I have very good preview results for windows 2 and 4 where DFL just build the given face. I also have some nice results in the last preview windows but most of them are still bad. sometimes the chin of the result is the other direction than the dst face, sometims its just a whirled messed up pixelation. in those cases i think it is because I dont have the same angle in my src.
i also have the problem that the result is not looking in the same directions as the dst face (eyes). skin color from src is taken but I didnt experiment with color options yet. Some dst faces have head down and some hair in front of the face, landmarks are ok, should I deleted those?
I have like 163k its now with his model and I wonder if I can add new src faces with missing angles to fix some problems? I also thinking about if I should just delete those dst faces with misaligned landmarks that I can not correct? I wonder if I do this if I have to retrain?
I also wonder if LIAF would be the better choice than DF for my project or if I messed it up by using some options in the start and/or left out later? in terms of loss i am at [0.3038][0.1437] should i wait train more or will less loss never fix angle/eye problems?
regards
edit: attached some examples. Here you see dst looking down getting filmed from above:
https://ibb.co/DwMn313
tried to fix chin area which has bad landmarks obv but wasnt able to
that obv results in a mess:
https://ibb.co/Hxp00Wj
another one where chin is other direction messing up everything:
https://ibb.co/2dDWSgX
on these problem I thought it is just slightly from down angle and I dont have that in src faces but still I have much faces looking in the left direction so it should make it?!
side profiles are very good btw. dst has bigger nose cheeks and chin but DFL manages that to be src shape just perfect.
edit2: one more question: when manually setting the landscapes should the blue bottom edge end one the real chin or the double chin when dst shows a double chin? do I have to retrain from scratch or can i redo dst files and train mistrained model?
I have to say: I want to make a deepfake with model src and pornstar dst and body simmilarity is most important so im limited in choosing the right face structre. but my dst pornstar is pretty close, the upper half of the face is similar to dst, the lower half is bigger, bigger chin, bigger mouth, to mention that
now I went by the old guide in the DFL 1.0 outdated thread 'My Personal DeepFake Workflow Using DeepFaceLab' from dpfks. I already went hard on sorting just the best src images, min res. 720p and only sharp images without shadows etc I think I did a pretty good job on this one.
the dst was harder for me because I have angles on the pornstar where camera is looking on the head pointing on the forehead or very from bottm pointing at the (double)chin. landmarks on these were faulty so I re extracetd manually but even here some of those are simply not correct landmarked no matter how I scroll or move the mouse. I tried the nearest and saved it.
I need to mention: I have very much angles of the src model from interview clips and such, but getting those very up and bottom angles is very hard to find so actually I dont have those very unusual angles in src. first question here is: If I find some how can I add them to my src and continue training without restart the whole model?
Now for settings atm I am using:
Code:
================= Model Summary =================
== ==
== Model name: model2porn_SAEHD ==
== ==
== Current iteration: 123507 ==
== ==
==--------------- Model Options ---------------==
== ==
== resolution: 128 ==
== face_type: f ==
== models_opt_on_gpu: True ==
== archi: dfhd ==
== ae_dims: 512 ==
== e_dims: 64 ==
== d_dims: 48 ==
== d_mask_dims: 16 ==
== learn_mask: False ==
== lr_dropout: False ==
== random_warp: True ==
== gan_power: 1.0 ==
== true_face_power: 0.01 ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: none ==
== clipgrad: True ==
== pretrain: False ==
== autobackup: False ==
== write_preview_history: True ==
== target_iter: 0 ==
== random_flip: False ==
== batch_size: 12 ==
== ==
==---------------- Running On -----------------==
== ==
== Device index: 0 ==
== Name: GeForce GTX 1080 Ti ==
== VRAM: 11.00GB ==
== ==
=================================================
at start I had batch size 8, unfortunately no clipgrad enabled, gan power 0, trueface 0, face style 10 bg style 10. changed at about 80k its like mentioned in the old HowTo. Now I have very good preview results for windows 2 and 4 where DFL just build the given face. I also have some nice results in the last preview windows but most of them are still bad. sometimes the chin of the result is the other direction than the dst face, sometims its just a whirled messed up pixelation. in those cases i think it is because I dont have the same angle in my src.
i also have the problem that the result is not looking in the same directions as the dst face (eyes). skin color from src is taken but I didnt experiment with color options yet. Some dst faces have head down and some hair in front of the face, landmarks are ok, should I deleted those?
I have like 163k its now with his model and I wonder if I can add new src faces with missing angles to fix some problems? I also thinking about if I should just delete those dst faces with misaligned landmarks that I can not correct? I wonder if I do this if I have to retrain?
I also wonder if LIAF would be the better choice than DF for my project or if I messed it up by using some options in the start and/or left out later? in terms of loss i am at [0.3038][0.1437] should i wait train more or will less loss never fix angle/eye problems?
regards
edit: attached some examples. Here you see dst looking down getting filmed from above:
https://ibb.co/DwMn313
tried to fix chin area which has bad landmarks obv but wasnt able to
that obv results in a mess:
https://ibb.co/Hxp00Wj
another one where chin is other direction messing up everything:
https://ibb.co/2dDWSgX
on these problem I thought it is just slightly from down angle and I dont have that in src faces but still I have much faces looking in the left direction so it should make it?!
side profiles are very good btw. dst has bigger nose cheeks and chin but DFL manages that to be src shape just perfect.
edit2: one more question: when manually setting the landscapes should the blue bottom edge end one the real chin or the double chin when dst shows a double chin? do I have to retrain from scratch or can i redo dst files and train mistrained model?