A little question

[angel-yi]

Hello, thank you very much for sharing the project I pulled the latest code and trained it according to your training steps I found that The current size of the image is 384, not 288 The syncnet and wav2lip models have v2 versions, and I couldn't find any relevant documentation to explain the differences between these two models. I am training HQ Wav2lip Sam_ When training. py, the import model was modified and it was found that it could not be trained The image I trained according to 384 has many different colored light spots on the edge or lip area of the image, but the overall resolution is relatively high and realistic hope

[kunyao2015]

same question

[einsqing]

How long have you been training?How effective is tooth restoration? @angel-yi

[angel-yi]

您训练多久了？牙齿修复效果如何？@angel-yi 我训练了几个小时，我的数据集不大，看到有光点效果就停了，牙齿就是下面这个效果

[angel-yi]

How long have you been training?How effective is tooth restoration? @angel-yi

I trained for a few hours, and my dataset is not large. When I saw the spot effect, it stopped. The teeth are like the following effect

[shahidmuneer]

You need a bit higher dataset and train the network until you get fake and real loss, for me after 30 thousand iterations. I am not getting the above spots.

[angel-yi]

You need a bit higher dataset and train the network until you get fake and real loss, for me after 30 thousand iterations. I am not getting the above spots.

Thank you, I am already increasing the size of the dataset and hope to improve this situation

[ghost]

you should FID metric to monitor which is best model and I think you should continue train when the loss is sideways

[lililuya]

Hi, I am a little confused about the size of 288 and 384, the train_syncnet_sam.py imports the "SyncNet_color_384", which is from the syncnet.py module. But the hq _wav2lip_sam_train.py imports the "SyncNet_color",which is from the syncnetv2.py module. I compare the two network and it's different. Is there i miss something important? I am looking for your reply! Thanks!

[ghost]

Hi you can check the syncnet should be 384, I'm too lazy to update so all the training should be in 384

[angel-yi]

you should FID metric to monitor which is best model and I think you should continue train when the loss is sideways

Thank you, I am constantly trying to train

[huangxin168]

Hi, I am a little confused about the size of 288 and 384, the train_syncnet_sam.py imports the "SyncNet_color_384", which is from the syncnet.py module. But the hq _wav2lip_sam_train.py imports the "SyncNet_color",which is from the syncnetv2.py module. I compare the two network and it's different. Is there i miss something important? I am looking for your reply! Thanks!

I also confused on this.

1. SyncNet_color_384 from the syncnet.py module
2. SyncNet_color from the syncnetv2.py module 1 and 2 are different modules

so in file hq _wav2lip_sam_train.py, should I modify this line: from models import SyncNet_color as SyncNet to from models import SyncNet_color_384 as SyncNet

[1059692261]

you should FID metric to monitor which is best model and I think you should continue train when the loss is sideways

Thank you, I am constantly trying to train

Have you achieve any success in training? I succeed in training syncnet, the loss has decreased to 0.25. However in wav2lip training, the sync loss remains at a large value around 4, and the sample images show that model just copy the mouth part of the reference image to the predict one.

[lililuya]

@huangxin168 Yes, u're right!

[kike-0304]

你应该使用 FID 指标来监控哪个是最好的模型，我认为当损失横向时你应该继续训练

谢谢，我一直在努力训练

您在培训中取得了成功吗？我成功训练了syncnet，损失减少到了0.25。然而，在 wav2lip 训练中，同步损失仍然保持在 4 左右的较大值，并且样本图像显示模型只是将参考图像的嘴部复制到预测图像。

you should FID metric to monitor which is best model and I think you should continue train when the loss is sideways

Thank you, I am constantly trying to train

Have you achieve any success in training? I succeed in training syncnet, the loss has decreased to 0.25. However in wav2lip training, the sync loss remains at a large value around 4, and the sample images show that model just copy the mouth part of the reference image to the predict one.

I have the same problem. After training, the inferred image is the same as the original image, and it seems that the audio did not play a role in driving the mouth.