Those finally come on there !! ControlNet tile for SDXL(&others)
Finally, "this" has arrived !!
It feels like I've been waiting for "this" for nearly a year... but finally, a "tile" ControlNet model that works with SDXL has arrived.
In addition to tile, SDXL models for basic functions like canny, openpose, and depth have also been released. There also seems to be a new feature called "union".
I'm already using it extensively in combination with A1111's MultiDiffusion Upscaler.


When ControlNet first supported SDXL, models for Cany and Openpose were released, but "tile" was the one most desired, yet no one had managed to develop it until now.
※10.7 2024 Postscript
Just as I was saying that, I learned that one had actually been developed back in March. I want to test this one as well.
As I mentioned in the article below previously, a similar model was developed by a Japanese creator, but it is not exactly the same as "SD15_tile". For starters, the intended use cases do not seem to be the same.
The biggest advantage of combining MultiDiffusion and SD15_tile in A1111 for upscaling is that it "dramatically improves detail density while maintaining the original image," but until now, that was basically impossible with SDXL.
Even looking at overseas forums, threads asking "Why is there no tile for SDXL?" were common.
In fact, I had been thinking the same thing all along.
That is precisely why the one that appeared in May 2024 below was, I think, groundbreaking.
ControlNet 852_a_clone_xl https://t.co/UKOp0ZBDRl
— 852話(hakoniwa) (@8co28) May 19, 2024
SDXLのCN 「clone」を公開しました。挙動としてはtileみたいなCNです。白黒画像に対してはprompt 「monochrome」で対応できます。
形状を維持するので、小さな画像を大きくしたり、AnimateDiff使用時の入力動画を維持するのに利用できます。 pic.twitter.com/s76wvQJPpg
However, while I think this model is excellent in terms of "maintaining the original image," which was one of the features of SD15_tile, it wasn't quite the same as SD15_tile in terms of "improving detail density," as it doesn't seem to have been created with that concept in mind.
The SDXL tile that was released this time is a massive 2.5GB on its own, but based on my testing in A1111 so far, it seems to have the performance to match.
t2t 1024×1024

i2i x2 upscaled

The SDXL tile released this time, along with the others like canny, openpose, and what appears to be the new union feature, are already attracting attention from overseas creators.
At the same time, thankfully, the Tiled VAE feature for SDXL has been restored in the recent update to the MultiDiffusion Upscaler for A1111.

This part has been a bit of a cat-and-mouse game for nearly a year, with Tiled VAE for SDXL becoming usable and then unusable repeatedly, but I am grateful that the MultiDiffusion Upscaler developer updates the program every time an error occurs.
