Can we train the model on a single video and can we add our own driving video

[vinay345]

trafficstars

[AliaksandrSiarohin]

Why you need to train on a single video?

[ExponentialML]

Yes, but the results will be satisfactory. I'm guessing that you're looking to do some kind of quick, supervised motion transfer and get around training the model as it takes a long time (I don't blame you, it's resource heavy). It's best that you create a dataset that you can train with the specific thing you want, and infer on that. In a scenario like this, it's possible you can create a synthetic dataset (for example, 3D animations of what you want where you can control the scene) then train on that.

@AliaksandrSiarohin I think the reason why is that it might be hard to build in the wild datasets for something you want. For example, if you wanted to train a model on people doing back flips, it would be hard to find sufficient data to train on due to the different camera angles. Datasets like Taichi are easier due to stationary cameras.

To answer the second question, yes.

[vinay345]

When i add my own driving video and the source image it doesnt work,what could be the reason Second thing can the driving video be anything other than the datasets videos that you have used

[AliaksandrSiarohin]

I can't say, show an example.

[Adorablepet]

When I use my own driving video and the original image, the faces in the generated video are different from those in the original image. I want to ask, what is causing this?The effect of generating video is not good. Thanks.

[AliaksandrSiarohin]

Send an example.

[Adorablepet]

@AliaksandrSiarohin test_demo.zip run command, as follows:

python demo.py --config config/ted384.yaml --driving_video driving_video.mp4 --source_image source_image.png --checkpoint checkpoints/ted384.pth

Thanks,

[AliaksandrSiarohin]

Your video is not cropped. You should crop it to make it square around the person. Plus you can try to use ted-youtube config and checkpoint.

[Adorablepet]

Your video is not cropped. You should crop it to make it square around the person. Plus you can try to use ted-youtube config and checkpoint.

I run command with ffmpeg:

ffmpeg -i driving_video.mp4 -vf scale=384:384,setdar=1:1 driving_video_crop.mp4

But result.mp4 is not so good. test_demo_crop.zip

Is there a problem with my cropping method?Could you give a reference code? Thanks again.

[zhaoloulou]

Thank for your great work! I'm going through the same thing，Is there something wrong with my data Desktop.zip

[zhaoloulou]

Hello, I would like to ask whether the model you provided is the best model and whether the training time and data are sufficient

[AliaksandrSiarohin]

Image and video crops should include upper part of the legs, see examples in Readme.