Is it possible to upscale output from custom graph/workflow in Krita ai diffusion?

[sfisher]

I have an existing SUPIR ComfyUI workflow that works pretty well for me and sometimes it might be convenient to use within Krita ai diffusion instead of running it in a separate ComfyUI web browser window.

The problem I'm running into is that it should increase the canvas size when upscaling. I used the Krita canvas input to the workflow and the Krita output widget at the end of the workflow.

It runs and creates larger output, but the output that krita ai diffusion seems to receive is just the top left corner of the image and the rest is cut off, seemingly cropped to the size of the original image canvas.

I thought maybe if I used the move layer tool or resized the canvas, I'd see the rest of the upscaled image because it was clipping visible output to the canvas size. But everything larger than the original canvas size seems to be discarded rather than just clipped to the canvas view area. At least I couldn't find a way to see the rest of the output that was generated by the custom workflow.

I didn't see other options or or widgets to output a different sized canvas than the original when I searched for "krita" widgets in comfyui. It would be nice to be able to get the rest of the image layer if the output image is larger than the input image canvas.

Am I missing a widget I should be using in ComfyUI or maybe this is a hard limitation of custom workflows right now that the output layer can't be larger than the canvas size.

It's not the end of the world since I can use ComfyUI separately and produce PNG files, but it would be nice to produce upscales through a custom workflow in Krita sometimes and I wasn't sure if I missed something.

ChatGPT tried to tell me to use the Krita Canvas widget and set the width and height, but it seemed to me that it was only an input widget for the workflow and not an output widget from it.

Thanks for such a great and feature filled plug in. I'd be willing to look at the code a bit to try to improve something if the changes were fairly minor and someone could point me to the areas I should look at. But I don't have a lot of background writing widgets for ComfyUI (but could likely figure it out some basics if it's not too complicated).

[Acly]

Result image is cropped at the moment (mostly to prevent crash when the image is smaller than canvas I think).

Here is a quick hack that might work in the way you expect (create a layer that is bigger than the canvas): b4ddf85

Haven't tested it, and it will certainly break some non-custom-workflow stuff, but might give you an idea where to look

[sfisher]

Thanks for the background. I added some additional code to detect upscaling (the output is at least 25% bigger in both dimensions) vs not upscaling. I think it's probably not a perfect way to tell it's being upscaled vs in-painted or something else, but it seems to work so far from the things I've tried.

main...sfisher:krita-ai-diffusion:custom-graph-upscale

When I just used the code without the crop by itself, it worked for upscaling, but when I did inpainting, it then didn't align the inpainted output layer correctly over the original layer anymore.

It seems some other minor things have changed from my local copy, but anyone who wants/needs this functionality could look at the apply_result method changes.

Basically the following code is different in that method:

image_width, image_height = image.extent.width, image.extent.height
bounds_width, bounds_height = params.bounds.extent.width, params.bounds.extent.height

is_upscaled = image_width > 1.25 * bounds_width and image_height > 1.25 * bounds_height

# if it's not upscaled it does the original crop
if not is_upscaled:
    image = Image.crop(image, Bounds(0, 0, *params.bounds.extent))

...
    # and the apply.behavior.replace ELSE branch

    if is_upscaled:
        self.layers.create(name, image, Bounds(0, 0, *image.extent))  # alternate behavior that preserves the full output
    else:
        self.layers.create(name, image, params.bounds)  # original behavior which outputs the cropped version

I don't know if this has an adverse affect on anything else, but it doesn't seem to from the things I've tried so far. (and assumes if you're upscaling it will be by more than 25% as the detection mechanism, which is sorta kludgy).