Tiling setting

[SirNate0]

Is it possible to have the script honor the Tiling setting for the image? The image itself seems to tile very nicely. However, the depth image has a distinct curve to it that does not tile. Using a high-pass filter after the fact mostly fixes it, but it is not perfect, leaving a slight seam.

With high-pass filter (and offset by 1/2 to show the seam)

[SirNate0]

While not seamless, if I turn off "BOOST (multi-resolution merging)" and turn on "Match net size to input size", the result does not have the broad sloping depth to the corner.

[SirNate0]

To actually make it seamless, it's possible to use the same approach as sd_hijack.py and simply force all the Conv2d layers to have the circular padding mode. E.g. adding this to load_models() in depthmap_generation.py, though I've only tested it on the res101 model (and it failed with BOOST enabled - probably have to make the same adjustment to the BOOST model, though it might be more complicated since I think it is breaking the image into different patches...)

    def apply_circular(enable):
        def flatten(el):
            flattened = [flatten(children) for children in el.children()]
            res = [el]
            for c in flattened:
                res += c
            return res
        layers = flatten(model)
        for layer in [layer for layer in layers if type(layer) == torch.nn.Conv2d]:
            layer.padding_mode = 'circular' if enable else 'zeros'
    apply_circular(True)

[semjon00]

Wow, this is impressive! I didn't even think of that, it indeed works without finetuning! A niche but very cool and creative usecase. I am adding this now.

[semjon00]

I wrote code that adds a new checkbox for the tiling mode, but my results are somewhat lacking...

[semjon00]

Ok, this feature has just hit the main branch. It still has some seams... It would be awesome if you could take a look (investigate 😄) and identify the reason...

[semjon00]

@SirNate0