run Style-Consistent Generation

[lllrf-xuexi]

I added eol, but the result is bad. How to use it

[zhuole1025]

Hi! I am not sure why you add eol for style consistent generation. The key of style consistent generation is to share the key and value within a batch of images for attention calculation.

[lllrf-xuexi]

I generate an image, W:H = 1024:512, so the latent is (1,4,64,128). When doing patchify_and_embed, I added eol_token in the middle and at the end of x, but the result is bad. Where else should I make changes?

[zhuole1025]

There is no eol token in Lumina-Next. You should treat each image at the batch dimension, so the rope will be added to each image independantly.