Support loading a batch of multiple images with LoadImage

[space-nuko]

You can specify a batch count for EmptyLatentImage. However LoadImage can only load one image. It would be nice to have batch support for that (even if not exposed to the frontend)

[space-nuko]

Also, ability to load one (or more) images and duplicate their latents into a batch, to be able to support img2img variants

[WASasquatch]

It needs a whole new file selection system to support this feature. Something that loads up the images in a selectable way.

But I don't know any init system that takes a batch of inits? Usually just one init for each batch gen.

My WAS Node Suite has a batch load node for directories, but it only works with prompt batching not batch gen.

[Zuellni]

Also, ability to load one (or more) images and duplicate their latents into a batch, to be able to support img2img variants

This is pretty simple, you just have to repeat the tensor along the batch dimension, I have a couple nodes for it. Loading multiple images seems hard.

[WASasquatch]

Also, ability to load one (or more) images and duplicate their latents into a batch, to be able to support img2img variants

This is pretty simple, you just have to repeat the tensor along the batch dimension, I have a couple nodes for it. Loading multiple images seems hard.

So you can do a batch run with batch inits by just giving a batched init? Curiously, and off-topic, do you know if ComfyUI supports torchvision transform tensors? Or is it different?

[Zuellni]

Yep, this works both with latents and images, and you can increase the batch size at any point in the workflow, not just when encoding.

I haven't done anything with torchvision so not entirely sure but looking at the docs it just uses a different shape. You should be able to work with the tensors if you first permute to 0, 3, 1, 2 and then back to 0, 2, 3, 1 at the end.

//I made a crude dir loading node that works with batch gens. The biggest issue (other than the lack of a picker) is that you can't torch stack the images if their dimensions are different (as expected) so they need to be resized/have padding added. Not sure if there's any way around that.