Change latent or output resolution of ksampler without cropping image.

[BrechtCorbeel]

I though latent was going to be a cool feature where an image would be an entity in abstract space that can be molded ignoring basic vectorised resolution in a sort of tensor like fashion, but apparently it is not. How do you make the ksampler interpret the image so I don't crop it and can set a resolution other then the input latent image?

Any guides and tutorials on latent nodes?

[ltdrdata]

In the context of SDXL, to put it simply and somewhat inaccurately, you can think of an 8x8 pixel image being converted into a single latent pixel.

When adjusting the aspect ratio without cropping, the latent values in the expanded direction must either be ignored and duplicated or interpolated like pixels.

To understand what happens in this process, you can check the effect by applying a Latent Upscale node.

The model used in KSampler is designed to predict noise, not to predict the empty spaces between latent values.

Custom nodes like NNLatentUpscale include a lightweight model trained for this purpose, but they are not a universal solution.

They are only effective for specific resolution conversions and only work with the latent representations used by certain models, since they were trained specifically for those latents.