The code of IBQ loss is mismatch with the paper eq10.

[yufan-aslp]

quant_loss = torch.mean((z_q - z)**2) + torch.mean((z_q_2.detach()-z)**2) + self.beta * \ torch.mean((z_q_2 - z.detach()) ** 2)

this code is at src/IBQ/modules/vqvae/quantize.py 453

But in the paper eq10 the beta is for the z with the z_q_2.detach()

[ShiFengyuan1999]

Hi @yufan-aslp, thanks for your reminder. Actually, we follow the implementation of the official VQGAN repo, leading to the same bug. We haven’t tried the case where legacy is set to False, but it should have a small impact on the final performance.

---