high-resolution image

[Lfg111]

Hello, I would like to ask how this model can handle high-resolution image pairs, such as those with a resolution of 4000*3000, while preserving as much image detail as possible?

[yyfz]

Hi, Lfg111!

Resolutions far beyond the training scope cause quality drops. You can check out training-free methods for positional encoding extrapolation to address this.

[Lfg111]

May I ask, what are the methods for extrapolating positional encoding without training?

[yyfz]

May I ask, what are the methods for extrapolating positional encoding without training?

Similar to this, but I am not sure if it works for discriminative models.