AI anon here, since think I actually discovered something interesting about the comic itself from my messing around.
I was looking to train a higher resolution Annie model using a technique that's worked on other webcomics, but every time I tried it always gave these ghastly results where the lines were all blurred and the colors smeared together. It was only happening with Gunnerkrigg Court too, so I investigated and when you zoom in on the comic panels they all have this odd smeared look that other webcomics don't have. Sort of like spreading a layer of Vaseline across them. Does anyone know what the cause of this is? Does Tom use some old resizing software to scale down his comic pages?
It's an old comic and the pages were uploaded at a really low resolution and he never increased it. Back in the day, it wasn't very low res, but nowadays it's tiny. Coupled with the texture he puts on everything and what seems to be not even the highest quality jpeg compression, as you can see here, in comparison to a jpeg that was compiled at max quality, it tends to be pretty blurry. Nut sure what you mean by 'smear' Maybe provide an example?
Sure, this one was intentionally scaled up to show it off but if you zoom in on this panel it has pretty much the same look. It looks like you're looking at the comic panels through a thin layer of film that's making the lines and colors sort of blend together. I know that there are some scans of the book releases and those look fine, but the comic pages universally have this issue and the LORA trainer can't seem to handle without giving really nightmarish results.
I guess that's true, it's a damn shame though. My issue is just that it means there's basically no way to get a LORA with Annie's modern look until we get scans of books that are up to this point in the comic, which is a shame because other webcomic characters work fine with the new methods and look wayyyy better than the shit I was posting before.
I was going to until I ran into this issue with the training images, now I legitimately don't think it's possible to get decent quality ones at the moment. Maybe in another year or two they'll have worked out a way so that the trainers are able to negate jpg artefact issues. The results I got trying to train on my old data sets were downright unsettling.
Basically the training model iterates upon minor details to try and replicate something to the maximal degree. If every training image has jpg artifacts it'll try and work those into the model itself, and it results in some odd stuff.
For comparison, here's an image I got from training on panels from Gunnerkrigg Court. Really shows the issues of expounding upon artefacts in the dataset, look at how weird and blurry everything looks. Now for another webcomic...
...Ava's Demon, also trained on the comic panels. Identical prompt, identical settings. Outside of some minor general AI wonkiness it's a world of difference. For reference that was the singular best result I got out of the Annie model, on average they were much worse.
[...]
A few anons have uploaded scans of the print books, they're in much higher resolution than what gets uploaded to the site. That might work better for training
Oh shit really? Is there a link to those?
1 month ago
Anonymous
interesting would love to a model trained on the Cummoner webcomic.
The comic has a rather simple artstyle, maybe some preprocessing to trace lines and fill with flat colours could help?
1 month ago
Anonymous
You'd basically need a way to retrace and recolor every single sample image in order to do away with the jpg artefacts, otherwise they're just going to end up being hard-baked into the model regardless. Not sure if there's a program that can do that automatically and I'm not nearly enough of an artist to be able to go in and do that by hand.
...Ava's Demon, also trained on the comic panels. Identical prompt, identical settings. Outside of some minor general AI wonkiness it's a world of difference. For reference that was the singular best result I got out of the Annie model, on average they were much worse.
[...]
Oh shit really? Is there a link to those?
>Same prompts
Have you tried putting something like jpeg artifacts or compression artifacts or jpeg noise or whatever in the negative prompts?
1 month ago
Anonymous
Yes, it doesn't have any notable effect because the model itself doesn't read those as jpeg artefacts, it just sees them as part of the training. Even if you specifically tagged each image as "jpeg artefacts/noise/etc." on the images it'd be useless because all of the images have that, so it'd just read it as an activation tag for the prompt. It's a case where you really just need cleaner samples.
AI anon here, since think I actually discovered something interesting about the comic itself from my messing around.
I was looking to train a higher resolution Annie model using a technique that's worked on other webcomics, but every time I tried it always gave these ghastly results where the lines were all blurred and the colors smeared together. It was only happening with Gunnerkrigg Court too, so I investigated and when you zoom in on the comic panels they all have this odd smeared look that other webcomics don't have. Sort of like spreading a layer of Vaseline across them. Does anyone know what the cause of this is? Does Tom use some old resizing software to scale down his comic pages?
A few anons have uploaded scans of the print books, they're in much higher resolution than what gets uploaded to the site. That might work better for training
>layer of Vaseline
He applies a texture over each page before publishing on the web (downscaled) or in print. The originals are in high resolution but also with this texture. Some of his twitch streams demonstrate this process and occasionally show the pre-textured layers. I've got the early volumes in hires digital but the last HD was volume 5.
...Ava's Demon, also trained on the comic panels. Identical prompt, identical settings. Outside of some minor general AI wonkiness it's a world of difference. For reference that was the singular best result I got out of the Annie model, on average they were much worse.
Holy based, thank you so much! This level of detail will definitely negate the issues with artefacts, it's a shame we can't quite get the modern style to work but hey, at least we can get a more classic look.
I don't have much experience training Loras but I know enough about stable diffusion that I highly doubt the subtle texture tom puts onto the page is going to appreciably impact the lora. Models are already trained on noisy data just fine.
It's actually a big problem when you're working with small images where the noise makes up a large part of the detail. Larger resolutions don't matter so much but samples from Gunnerkrigg's website will definitely be small for obvious reasons.
If you search "Read Gunnerkrigg Court online" on google you can find several websites with scans up to book 5. They stop there for some reason though, even though we're up to book 8. It's the older style but much cleaner, and you should be able to get a pretty good Annie out of them if you start with book 3.
Thanks anon, I'll give that a try. Currently training another LORA and I'll have to make a whole new dataset from scratch, so this thread might be dead by the time I have some results, but I'll post the results of this once it's all done.
Place your bets Cinemaphile. Do you think Tom will actually make Omega an interesting character? Or at least charming like Lana?
Or will she be lame and boring?
Tom is horrible with payoffs but he can still introduce cool characters... right?!
i mean it's already pretty interesting. like is omega some kinda skull with robot stuff going into it? what's that about, is it the opposite of the new humans, like the court turned a human brain into a cpu? this black-haired girl Omega could be a zimmingham illusion or coyote forgetting that he's not Omega or something
Looking sort of like Zimmy shouldn’t be a coincidence. There is probably a connection between Omega, Zeta, and Gamma. Someone is a clone or a tulpa of someone else.
>Having a chapter called Omega that isn't the last chapter of the comic
That's forbidden!
>Inb4 it's revealed that Annie's real name is not Antimony, but Alpha.
>Alpha and Omega
TOOOMM!!
But Annie was named after Surma's grandmother. She wanted her child to be named Antimony and Tony calls her Antimony, too, anon.
And you think current Tom wouldn't do a rercon?
Yes? Obviously?
Tom can do the funniest thing by bringing it all to a sudden end.
So that's it? Tom strikes again? Zaloupoyota ends up being another Plain Jane instead of almost /d/-tier local drawings?
We haven't seen Zimmy, Loup, Coyote or Lana yet.
>Zaloupoyota
Is that some Black person deity?
gunnerblart mall cop
Paul Blart is a much more competent protagonist than Annie.
5$ on Omega wanting the same freedom the robots got because frick the court and their plots
AI anon here, since think I actually discovered something interesting about the comic itself from my messing around.
I was looking to train a higher resolution Annie model using a technique that's worked on other webcomics, but every time I tried it always gave these ghastly results where the lines were all blurred and the colors smeared together. It was only happening with Gunnerkrigg Court too, so I investigated and when you zoom in on the comic panels they all have this odd smeared look that other webcomics don't have. Sort of like spreading a layer of Vaseline across them. Does anyone know what the cause of this is? Does Tom use some old resizing software to scale down his comic pages?
It's an old comic and the pages were uploaded at a really low resolution and he never increased it. Back in the day, it wasn't very low res, but nowadays it's tiny. Coupled with the texture he puts on everything and what seems to be not even the highest quality jpeg compression, as you can see here, in comparison to a jpeg that was compiled at max quality, it tends to be pretty blurry. Nut sure what you mean by 'smear' Maybe provide an example?
Sure, this one was intentionally scaled up to show it off but if you zoom in on this panel it has pretty much the same look. It looks like you're looking at the comic panels through a thin layer of film that's making the lines and colors sort of blend together. I know that there are some scans of the book releases and those look fine, but the comic pages universally have this issue and the LORA trainer can't seem to handle without giving really nightmarish results.
Well, yeah, it's just low resolution and lossy compression. I don't think anything weird is going on with these jpgs though.
I guess that's true, it's a damn shame though. My issue is just that it means there's basically no way to get a LORA with Annie's modern look until we get scans of books that are up to this point in the comic, which is a shame because other webcomic characters work fine with the new methods and look wayyyy better than the shit I was posting before.
Hi AI anon! I miss your stuff. Are you gonna do more Gunnerkrigg AI pics?
I was going to until I ran into this issue with the training images, now I legitimately don't think it's possible to get decent quality ones at the moment. Maybe in another year or two they'll have worked out a way so that the trainers are able to negate jpg artefact issues. The results I got trying to train on my old data sets were downright unsettling.
Or I guess if high quality scans of the more recent books became available, but so far as I'm aware those aren't really a thing.
Aww shucks
I don't know much about AI, but why would the compression mess with it that much?
Basically the training model iterates upon minor details to try and replicate something to the maximal degree. If every training image has jpg artifacts it'll try and work those into the model itself, and it results in some odd stuff.
For comparison, here's an image I got from training on panels from Gunnerkrigg Court. Really shows the issues of expounding upon artefacts in the dataset, look at how weird and blurry everything looks. Now for another webcomic...
...Ava's Demon, also trained on the comic panels. Identical prompt, identical settings. Outside of some minor general AI wonkiness it's a world of difference. For reference that was the singular best result I got out of the Annie model, on average they were much worse.
Oh shit really? Is there a link to those?
interesting would love to a model trained on the Cummoner webcomic.
The comic has a rather simple artstyle, maybe some preprocessing to trace lines and fill with flat colours could help?
You'd basically need a way to retrace and recolor every single sample image in order to do away with the jpg artefacts, otherwise they're just going to end up being hard-baked into the model regardless. Not sure if there's a program that can do that automatically and I'm not nearly enough of an artist to be able to go in and do that by hand.
>Same prompts
Have you tried putting something like jpeg artifacts or compression artifacts or jpeg noise or whatever in the negative prompts?
Yes, it doesn't have any notable effect because the model itself doesn't read those as jpeg artefacts, it just sees them as part of the training. Even if you specifically tagged each image as "jpeg artefacts/noise/etc." on the images it'd be useless because all of the images have that, so it'd just read it as an activation tag for the prompt. It's a case where you really just need cleaner samples.
Probably because such noise is predictable, so the model learns that pattern as a part of the image
A few anons have uploaded scans of the print books, they're in much higher resolution than what gets uploaded to the site. That might work better for training
>layer of Vaseline
He applies a texture over each page before publishing on the web (downscaled) or in print. The originals are in high resolution but also with this texture. Some of his twitch streams demonstrate this process and occasionally show the pre-textured layers. I've got the early volumes in hires digital but the last HD was volume 5.
>link
libgen
I don't think this would be an issue at the resolution you posted. Also I'm not finding it on libgen
libgen.gs search for gunnerkrigg, page 3 of listings; I see volumes 5 and 6 now also; but I bought 1 to 4 via humblebundle, picrel example.
Holy based, thank you so much! This level of detail will definitely negate the issues with artefacts, it's a shame we can't quite get the modern style to work but hey, at least we can get a more classic look.
Actually I might just be a dum-dum but none of the download links seem to work? Even the torrents seem to be invalid.
Scratch that, figured it out and I am indeed a dum-dum.
I don't have much experience training Loras but I know enough about stable diffusion that I highly doubt the subtle texture tom puts onto the page is going to appreciably impact the lora. Models are already trained on noisy data just fine.
It's actually a big problem when you're working with small images where the noise makes up a large part of the detail. Larger resolutions don't matter so much but samples from Gunnerkrigg's website will definitely be small for obvious reasons.
If you search "Read Gunnerkrigg Court online" on google you can find several websites with scans up to book 5. They stop there for some reason though, even though we're up to book 8. It's the older style but much cleaner, and you should be able to get a pretty good Annie out of them if you start with book 3.
Thanks anon, I'll give that a try. Currently training another LORA and I'll have to make a whole new dataset from scratch, so this thread might be dead by the time I have some results, but I'll post the results of this once it's all done.
Place your bets Cinemaphile. Do you think Tom will actually make Omega an interesting character? Or at least charming like Lana?
Or will she be lame and boring?
Tom is horrible with payoffs but he can still introduce cool characters... right?!
Omega will be very cool and fun!
Omega will be very bland and boring!
Omega will be a plot device and not much more.
Exactly! And on top of that she is also (so far) boring looking and unappealing.
Nah she's cute, I would.
The human form is kinda boring though.
Imagine what she can do with those eyesockets bro!
i mean it's already pretty interesting. like is omega some kinda skull with robot stuff going into it? what's that about, is it the opposite of the new humans, like the court turned a human brain into a cpu? this black-haired girl Omega could be a zimmingham illusion or coyote forgetting that he's not Omega or something
Looking sort of like Zimmy shouldn’t be a coincidence. There is probably a connection between Omega, Zeta, and Gamma. Someone is a clone or a tulpa of someone else.
It's probably just the remains of a person with very strong predictive abilities who is being kept alive through technomagic.
I installed an ai because I wanted to give it a shot.
It tried. Pls don't bully.
>no amputation
Boooooo!!
It's not bad, but obviously it has problems. the shitty AI image couldn't even replicate the hairband...
>Kat being girly and enjoying it
Blasphemy!
Omega's cute eyesockets!
But can Omega see why kids love the taste of Cinnamon Toast Crunch?