search

LTH14 / rcg

PyTorch implementation of RCG https://arxiv.org/abs/2312.03701
MIT License
785 stars 36 forks source link

Question about figure7, could you please introduce your interpolation algorithm? #20

Open ZCG335 opened 8 months ago

ZCG335 commented 8 months ago

As there is no code related to figure7 interpolation in the GitHub repository.

LTH14 commented 8 months ago

The feature f is just a linear interpolation f=t*f1 + (1-t)*f2 of two features f1 and f2 extracted from two images. After linearly interpolating the features, we use the pixel generator to generate the image conditioned on f.

ZCG335 commented 8 months ago

Thank you very much! I have run the interpolation code successfully.

mapengsen commented 4 months ago

@LTH14 Dear author, I would like to ask about Fig7, do you directly add up the rep of the two images and input the interpolation rep result obtained by pixel generation?

In the implementation of the code is: interpolation_rep = 0.1 rep1 + 0.9 rep2

thank you very much!

LTH14 commented 4 months ago

For Figure 7, we first extract the MoCo-v3 representations from the two real images (rep1 and rep2), and then obtain the interpolated representation as inter_rep = trep1 + (1-t)rep2. Then the image pixel generator will condition on this inter_rep to generate the image with interpolated semantics.

mapengsen commented 4 months ago

For Figure 7, we first extract the MoCo-v3 representations from the two real images (rep1 and rep2), and then obtain the interpolated representation as inter_rep = trep1 + (1-t)rep2. Then the image pixel generator will condition on this inter_rep to generate the image with interpolated semantics.

Thank you for your kind reply. I think I understand