DeepFaceLab
DeepFaceLab copied to clipboard
iperov
how long train ?
how long does deepfacelab have to train? 1 hour ? 4 hours ? 12 hours ? 1 day ?
It depends on how much quality you want for the fake. Also strongly depends on the model settings, eg the Dimensions, batch size and resolution. I usually train at least 100k with SAEHD, which takes roughly 100.000x1.0 sec = 27 hours (on Core i7 and RTX2060S) , then do a merge to see whether the quality is good enough. If not, then more train is needed. You can also always monitor the quality in the training preview.
But I think these topics are better discussed in the forums mentioned under Links on https://github.com/iperov/DeepFaceLab
Es hängt davon ab, wie viel Qualität Sie für die Fälschung wünschen. Hängt auch stark von den Modelleinstellungen ab, z. B. den Abmessungen, der Stapelgröße und der Auflösung. Normalerweise trainiere ich mindestens 100.000 mit SAEHD, was ungefähr 100.000 x 1,0 Sekunden = 27 Stunden dauert (auf Core i7 und RTX2060S), und führe dann eine Zusammenführung durch, um festzustellen, ob die Qualität gut genug ist. Wenn nicht, wird mehr Zug benötigt. Sie können die Qualität auch jederzeit in der Trainingsvorschau überwachen.
Ich denke jedoch, dass diese Themen in den unter Links unter https://github.com/iperov/DeepFaceLab genannten Foren besser behandelt werden
I have trained now I want to convert the trained model but in the directory there is no CONVERT . BAT , but MERGED to MP4. Asked. and when I run MERGED to MP4.bat, it does not export VIDEO but only a PHOTO with audio TON, in video format. however, the VIDEO does not move.
I have trained now I want to convert the trained model but in the directory there is no CONVERT . BAT , but MERGED to MP4. Asked. and when I run MERGED to MP4.bat, it does not export VIDEO but only a PHOTO with audio TON, in video format. however, the VIDEO does not move.
... Step 6) trainer Step 7) merger --> The convertor you are looking for, it is renamed. Step 8) merged to mp4 --> Final step to get result video. is is not a convertor.
What I dont understand is why the training selects the CPU as device, and not the GPU available.
Do you have to train some model, for each deepfake you make, or can you train once for any source/target you put in
