generate images with arbitrary resolutions,

[Leiii-Cao]

is there any way to use a checkpoint from that model to generate images with arbitrary resolutions,

[LTH14]

You can consider using similar techniques as MaskGIT: first, generate a 256x256 image, and keep expanding it to meet the resolution requirement.

[Leiii-Cao]

您可以考虑使用与MaskGIT类似的技术：首先，生成一个256x256的图像，并不断扩大它以满足分辨率要求。

I roughly understand your approach, but if I want to apply it to downstream tasks like image editing, wouldn’t the resolution of the edited images be limited, and wouldn’t it be challenging to work with arbitrary resolutions such as 345x566?

[LTH14]

Since we use a learnable position embedding, it is hard to transfer it to different resolutions. If you want to generalize to different resolutions, you could consider using ROPE as the position embedding during training.

[Leiii-Cao]

由于我们使用可学习的位置嵌入，因此很难将其迁移到不同的分辨率。如果您想要推广到不同的分辨率，可以考虑在训练期间使用 ROPE 作为位置嵌入。

Thank you for your explanation. Your answer was very helpful to me.