ControlNet-v1-1-nightly
ControlNet-v1-1-nightly copied to clipboard
lllyasviel
Tile?
how to use TILE? it is super interesting but i dont get it, tryed everything
I'm in the same boat. I got what seemed like it working on img2img with tiling turned on and Ultimate SD Upscaler turned on, but the image kinda took 4 hours on my card and looked like crap. Nothing was consistent and most of it ended up with a bunch of aberrant noise.
I would love to see some more in depth instructions on Tile. I have figured out several of the other control nets on my own, but this one just feels off. Also, putting the image in as a setting and/or setting the tile-gaussian pre-processor breaks the whole thing entirely.
it is not finished yet, we are thinking about better ways to make use of it
It seems that after an update tile_gaussian is gone, and now there is only tile_resample. Is it possible to get it back somehow?
It seems that after an update tile_gaussian is gone, and now there is only tile_resample. Is it possible to get it back somehow?
tile_gaussian is not needed now. the new tile model v11f1e_tile has learned transforms inside neural networks so that you do not need tile_gaussian to blur the image using preprocessors. note that the new tile model v11f1e_tile can receive none as preprocessor.
In any case if you strongly prefer the previously unfinished tile model, you can simulate the effect of previous tile_gaussian with photoshop's Gaussian Blur (Filter-Blur-Gaussian Blur). and then set none as preprocessor to use the previous unfinished tile model. tile_gaussian is essentially a Gaussian Blur.
Note that we plan to not have tile_gaussian because tile_gaussian is not the correct preprocessor for v11f1e_tile, and including it may cause many confusion. Again, directly taking original image as input is much more technically correct - v11f1e_tile can directly take original image as input (using preprocessor none)
The previous unfinished tile model has a tag [u], meaning that it may be updated when finished. Note that we have long-term support for all production-ready [p] and experimental [e] models. An evidence is that we did not remove "normal_midas" even when "normal_bae" is a much more correct method. However, the previous tile model has the tag [u], and we cannot offer long-term support for a unfinished model since that model has too many problems.
Oh, I think I understand, so it has now reached a state where it is more advanced, so the blurring isn't need anymore.
Thank you, the confusion was - I think - that I used the previous tile model. And I didn't know that it has been updated now. How can I know that in the future, is the a place where I can sign up for notification about these?
Thank you that you took the time to answer my question, and I appreciate your work.
Oh, I think I understand, so it has now reached a state where it is more advanced, so the blurring isn't need anymore.
Thank you, the confusion was - I think - that I used the previous tile model. And I didn't know that it has been updated now. How can I know that in the future, is the a place where I can sign up for notification about these?
Thank you that you took the time to answer my question, and I appreciate your work.
We are research lab and do not have enough human resources or efforts to promote or advertise. But it seems that we will have a official online documentation soon. The SD community does not seems to have many documentations. Even A1111 does not have a documentation.
Oh, sorry for the confusion, I thought about a github watch - so what to watch and where so these wont be missed.
Could you share some details of training the tile controlnet, e.g. tile pairs
@lllyasviel Could you share some details of training the tile controlnet and shuffle controlnet, e.g. the training dataset and ticks
