Closed XiaoHaoPan closed 10 months ago
Hi, @XiaoHaoPan,
Thank you for answering, I have another question, please ask the video generated in the first phase of the test, where the second frame generated the frame, the reference frame used is the first real driver frame, or the model generated fake driver frame?
Please ask how the settings in your paper are different from those in the code. Do I need to change the epoch settings to achieve the pre-training effect you provide?
Hi, @XiaoHaoPan, during testing, we use DM to generate a sequence of flow maps and apply flow maps to the given real reference image to generate a sequence of generated frames. We always keep warping the real reference frame to maintain good quality.
Thanks!I looked at your training DM phase code and I found that these two settings are different for different datasets. Is it because of the differences in different datasets?
MUG:
MHAD:
NATOPS:
Did the pre-trained model you provided train with this setup?
Hi, @XiaoHaoPan, Sorry for misunderstanding your second question! I have updated my previous comments. Yes, the size of NATOPS is much larger than the MUG and MHAD. So I set a smaller number of epochs. The pre-trained models I provided should be consistent with these settings if I didn't edit the code later. But it has been several months ago and I am not very sure about that.
1.What are the three losses in the first stage?
2.The generator model in this parameter seems to contain three trainable networks in one stage, is it correct?
3.What is the purpose of adding x0 in the second stage, I did not find it in the paper
4.I am not very clear about the division of test sets and datasets