Compare to SimVQ

[suruoxi]

Have you compared IBQ to SimVQ? I think SimVQ has the similar motivation of "optimize all codebook embeddings".

[ShiFengyuan1999]

Hi @suruoxi, thanks for your interest in our work.

SimVQ has a similar motivation to IBQ, but we adopt different methods to optimize all codebook embeddings. SimVQ optimizes a linear transformation space $W \in R^{D \times D}$, while our IBQ optimizes the codes ($C \in R^K\times D$) themselves, where $K\gg D$.

It can not compare SimVQ with IBQ directly, because SimVQ is trained on ImageNet 128x128 and IBQ is trained on ImageNet 256x256. Moreover, SimVQ only shows the reconstruction performance and does not give the generation results. But we can provide some results from the papers for reference.

Method	Train Resolution	Codebook Size	rFID $\downarrow$
SimVQ	128 $\times$ 128	262144	1.99
OpenMAGVIT2	128 $\times$ 128	262144	1.18
OpenMAGVIT2	256 $\times$ 256	262144	1.17
IBQ	256 $\times$ 256	262144	1.00

[suruoxi]

thanks for the reply, the rFID of IBQ is impressive.