I addressed all comments. The public APIs of OMEZarrWriter and PyramidalOMEZarrWriter are now similar (this required some API changes on OMEZarrWriter). Here is a script on how to use them:

import qupath.lib.images.writers.ome.zarr.OMEZarrWriter
import qupath.lib.images.writers.ome.zarr.PyramidalOMEZarrWriter

var path = "/path/to/img.ome.zarr"
var server = getCurrentServer()

new OMEZarrWriter.Builder(server)   // or new PyramidalOMEZarrWriter.Builder(server)
    .downsamples(1, 4, 16)
    .build(path)
    .writeImage()

It is now possible to provide downsamples without including 1.

Thanks, I think there's a bug in PyramidalOMEZarrWriter (I couldn't replicate it with OMEZarrWriter) connected to tile caching.

I ran the following code for 2 different images (you'll recognize them...), deleting the zarr in between:

import qupath.lib.images.writers.ome.zarr.OMEZarrWriter
import qupath.lib.images.writers.ome.zarr.PyramidalOMEZarrWriter

var server = getCurrentServer()
var path = buildPathInProject("anything.ome.zarr")

new PyramidalOMEZarrWriter.Builder(server)
    .downsamples(16, 64)
    .build(path)
    .writeImage()

It produced an image like this:

I initially thought that it was using the full image width and height, but those do in fact seem correctly downsampled (7936 x 4608px) - and the image displays properly when zoomed in so that pixels are requested from the highest available resolution.

See also #2012 - it's a longstanding bug, and perhaps fixing it would be enough.

I fixed a bug occurring when the first requested downsample isn't one.

That doesn't fix your issue though. When I run your workflow, I also get a bit a CMU-1 in the output image. However, if I clear the tile cache between the two writes, then the output image is correct. So, I think this is clearly linked with #2012.

How do you think we should fix that? Automatically clear the cache if two consecutive writes to the same location are detected?

Please check out #2013 and see if that fixes the issue - I'm not sure how exactly the modified time behaves with directories.

If you're temporarily creating a BioFormatsImageServer, then I think it would make sense to clear the cache specifically for that when it's serves its purpose - although I'm not sure how easy it is to get access to the required objects to call that method.

#2013 fixes the issue.

I'll merge both then, thanks!