seems to be some tiling going on with the non dev model
my testing is showing some tiling on the outputs. the dev model is fine so far.
2048x2048
cfg- 5.0
shift - 3.0
steps - 50
These nodes makes the grids way less noticable and makes the colours way more natural.
These nodes makes the grids way less noticable and makes the colours way more natural.
sadly, they cant be fed into my nodes in the current state their in, different output from the model node than the string allows
It's an issue in the model itself, even their example images have the same pattern, and I've ran their code alongside my implementation to confirm this too. Hope is that with all the tools available in ComfyUI we can better mitigate this, as the community is already discovering.
I love the model so far, but there are noticeable compression-like artifacts for me as well. I'm waiting for an update to Ostris AI Toolkit, and then I'll be able to see if I can find a way to train this behavior out of the model. I just hope it's caused by bad training data and not the model architecture itself, because I've been liking it otherwise.
Here's a sample where the blocky artifacts are especialy noticeable... it's quite insane to me.
Here's a sample where the blocky artifacts are especialy noticeable... it's quite insane to me.
Thats a very nice output tho all things considered. Great overall composition and i like the lighting
I've added a node that works around it a bit, basically just run extra passes with the patch grid offset near end of the sampling
Default settings already help:
With more passes (more offsets), slower and causes overall blur, but gets rid of the seams:
Cleanup nodes? Will they work on Ernie Image Turbo? Would be great.
Thats a very nice output tho all things considered. Great overall composition and i like the lighting
Thank you! The workflow might still be in the image if you want to take a look. I use a few tricks to control things like lighting, composition, setting, subject, etc etc.
I've added a node that works around it a bit, basically just run extra passes with the patch grid offset near end of the sampling
I've came to realize that training won't get rid of the patches, so I'm looking forwards to trying your fix!