[Bug] width or height can not be 608

[FNsi]

GGML_Assert ggml.c 5733 a->ne[0] .....

[leejet]

The height and width should be multiples of 64. I have updated the prompt for the command line parameters

[FNsi]

The height and width should be multiples of 64. I have updated the prompt for the command line parameters

Nice work🙏

[dmikey]

Shouldn't they be multiples of 8?

504x520 is a valid dimension in other SD pipelines.

[leejet]

Shouldn't they be multiples of 8?

504x520 is a valid dimension in other SD pipelines.

The sd diffusion process run in the latent space of 1/8 of the width and height. In addition, unet will downsample the middle layer to 1/8 during the operation, so in fact the middle layer size will be 1/64 of width and height, which is why sd.cpp requires width and height to be multiples of 64.

If "sd" were to scale the latent space to a multiple of 8 before U-Net computation and then restore it to its original size afterward, it would indeed allow input widths and heights to be multiples of 8. However, I have not implemented it yet.

[rootgrandfather]

Is there any plan to implement this feature?

[stduhpf]

Is there any plan to implement this feature?

This already works for Flux and SD3 models if you bypass the part that sanitizes the cli arguments.

For example, if you use the server example (#367) , you can set the width and height in the json payload, and the values aren't being checked. (this causes crashes with Unet-based models btw)

json payload : {'prompt':""" Two kittens (a black one and a white one) holding together a large sign that says "Flux.cpp works in 720p!", they are on either side of the sign. There are out of focus balls of wool in the background.""",'height':720, 'width':1280, 'seed':42})

Or, more relevant to the issue: