Questions about generation quality

[xizaoqu]

Hi, thanks for presenting this interesting paper.

Table 2 shows that Show-o achieves impressive generation ability (better than SDv1.5) with a much smaller training scale.

Could you provide some insights about why the discrete diffusion process can be even better than continuous diffusion with many denoising steps?

Thanks.

[Sierkinhane]

Hello, as the laion data is currently not accessible, it's a bit hard for a fair comparison. However, in our experiment, what we want to demonstrate is that with relatively sufficient data, discrete diffusion is able to have a good performance. I think one of the most important factors is the quality of image-text pairs.