What the "Controlnet tile" model really does and How was it trained?

[moonryul]

Hi,

I read that "controlnet tile" was trained by "image tiling" or on "tiled image". It suggests that the input condition to this model is a tiled image; But.... When I read gradio_tile.py, the "control" input is just a upsampled/resized image obtained from an input image. It has NOTHING to do with image tiling.

I know this model is used in conjunction with "Ultimate SD supscale" script in Automatic1111. But when I looked at the script, it splits the input image to tiles and invokes the controlNet for each tile, and combines the resulting tiles. So, tiling has to do with the script and nothing to do with the "controlenet-tile", it seems.

So, I wonder: 1) Q1. In what sense, is it a "controlnet-tile" model? 2) Q2. What kind of dataset is used to train this tile model? In the case of "inpaint" model, the dataset consists of pairs of (image, [mask, masked_image]). In the case of "upscaling" model, the dataset are pairs of (image, low_resolution version).

But in the case of "controlnet-tile" model, the dataset is not clear.

Sincerely Moon

[geroldmeisinger]

duplicate of #125 which was 5 lines down

[moonryul]

duplicate of #125 which was 5 lines down

Thanks for the answer. Do you mean the answers to my questions can be found in #125?

But I could not.