Variable Image size

[divamgupta]

Right now the trivial way to generate images of different sizes is to recreate the Maple Diffusion object with the new size params. Is there any way which is more efficient?

[madebyollin]

Per Apple's WWDC session / docs it should be possible to support variable image sizes efficiently:

1. Replace the height / width usage in UNet placeholders with -1 to indicate an unspecified dimension
2. Set a flag to skip type inference during UNet graph compilation

I started the code to do this in a branch here: f590ea8. Currently it fails to compile the graph, though.

I think it's probably going to be possible-but-annoying to get variable shapes working - since it requires editing the graph code in order to make sure everything compiles (this may require adding some graph.shapeOf operators when reshaping an hw dim into separate h and w dims, for example, or just never creating any hw dims).

[Gitterman69]

Diffusion Bee integrated the approach in a perfect way!!!

[ShirasawaSama]

@madebyollin Hello, is there any new progress on this issue?

[divamgupta]

I actually implemented it with the help of @madebyollin . Code is a bit messy I can make it public or make a PR some day

[ShirasawaSama]

This is good news, and I can also help you organize the code if you need it.