guessing multiscale is counter intuitive?

[HelenaLC]

• .guess_scale currently "rounds up to the next scale"
• e.g., width=height=11 for scales 20x20 and 10x10 would choose 20x20
• my intuition would be to instead minimize the total difference in input/output size
• e.g., w=h=11 would choose 10x10, w=h=16 20x20, and w=h=15 is for us to decide

Thumbs up if you agree @Artur-man, else happy to discuss/learn about the reasoning behind the current implementation!

[Artur-man]

ah very good point ... when I have discussed this with @melonora and @LucaMarconato, they mentioned to me that you first pick the scale that is closest (but higher) to the target resolution, from there, you rasterize and decrease the resolution to fit the targeted resolution.

Here is a good example:

This is the multiscale image of VisiumHD example (3, 21943, 23618), (3, 10971, 11809), (3, 5485, 5904), (3, 2742, 2952), (3, 1371, 1476)

Say that you wanna visualize the image with ... (2500, 2600) resolution. The closest resolution in the list of scales is (2742, 2952) then, if needed, you downgrade the resolution from (2742, 2952) to (2500, 2600), I thought it is better the change the resolution from a higher resolution rather than a lower resolution, meaning when you transform the (1371, 1476) image to (2500, 2600) you lose more information.

However, I am happy with which strategy you wanna go, at the end of the day the idea is to get a lower resolution image in the memory :D

[melonora]

Hello there! Please see https://github.com/scverse/spatialdata-plot/blob/0667ef8bf0633ad9574bbf4dd6e5defd60a6762f/src/spatialdata_plot/pl/utils.py#L1282C6-L1352 for the implementation in spatialdata-plot. This takes the figure width, height and dpi into account.

[melonora]

Thinking back I think for choosing the higher scale was mainly due to the fact that you would have to choose an interpolation so that would be another parameter.