A discussion about using AI to grade comic books. 3 3

[namisgr]

On 12/22/2023 at 10:12 PM, MattTheDuck said:

Not sure why this is the case. 

If someone can build one AI that can do this job, they can build a million of them.  Could probably get every book currently waiting at CGC graded in a lazy afternoon.

A computerized and robotic system could run around the clock.  Humans program upcoming activities, and load the hopper with books to be graded.

Encapsulation, labeling, and quality control also have the potential to be automated.

 

[Prince Namor]

And again, it wouldn't necessarily remove the possibility of ANY errors... but it could remove the possibility of someone getting paid under the table to gift grades.

CGC, outside of many of the other beneficial reasons for doing this, could also look at it as a protection from being taken advantage of by the riff raff that has preyed upon this hobby since... well, since it's been a hobby.

[BuscemasAvengers]

AI should be able to detect/quantify moisture damage, shouldn't it? Rusted staples/migration as well? Even signatures that haven't been witnessed?

[lizards2]

On 12/23/2023 at 3:03 AM, Get Marwood & I said:

 

Do you remember that high grade modern that was printed to look like it was a old beat up comic? CGC graded it a 2.0 or something like that and we all laughed thinking a person had made the error. Maybe no human ever looked at it.

 

I bought a raw Sea Devils a few years back from a boardie. Interestingly, he had assigned a much higher grade than the prominent company he had apparently purchased the book from. I was looking the book over, and aha - there was the reason the prominent company graded it so low, a big, long, hard crease. Upon further inspection, that flaw was on the original artwork. So, the boardie had it right and the company was probably bag-grading (which I am pretty sure is the case after subsequently buying some raws from them). I have a Magnus, Robot Fighter comic that has a similar flaw in the artwork, IIRC.

[Get Marwood & I]

On 12/23/2023 at 12:56 PM, BuscemasAvengers said:

AI should be able to detect/quantify moisture damage, shouldn't it? Rusted staples/migration as well? Even signatures that haven't been witnessed?

I don't see why not. There are many possibilities, I should imagine.

Would it bother you, to learn that AI was grading comics? I don't think I'm that fussed really - certainly not on moderns.

On 12/23/2023 at 2:28 PM, lizards2 said:

I bought a raw Sea Devils a few years back from a boardie. Interestingly, he had assigned a much higher grade than the prominent company he had apparently purchased the book from. I was looking the book over, and aha - there was the reason the prominent company graded it so low, a big, long, hard crease. Upon further inspection, that flaw was on the original artwork. So, the boardie had it right and the company was probably bag-grading (which I am pretty sure is the case after subsequently buying some raws from them). I have a Magnus, Robot Fighter comic that has a similar flaw in the artwork, IIRC.

Maybe robots are just as useless as people Liz. Or at least, the people that programmed them.

Didn't one of those driverless cars crash into the side of a lorry that had a picture of the sky on it once? Something like that. 

This one below could play for Man U.

 

[CGC Mike]

On 12/22/2023 at 9:35 PM, VintageComics said:

...so as not to distract from the other thread. 

Have at it! 

I locked the last thread you started about AI.  It's a violation of the terms of use to start another one after the original thread has been locked.  You should have asked me before starting this one.  I might have said ok.  

[CGC Mike]

I am going to unlock this thread, as I can see where it might be an important subject in the comics industry.  If it goes like the last one I locked, I will be vary generous with points to any and all accounts that I feel are a violation of the terms of use.  Everyone is supposed to be 18 years or older here.  I expect you to act like an adult.  No trolling or any contentious posts will be tolerated.

[Chip Cataldo]

Thanks Mike. That's nice of you to give Roy a Christmas present. 

[shadroch]

Someone correct me if I'm wrong, but I believe the technology already exists to scan a book without opening it.  It would be easy to note missing pieces, handwriting, etc. 

You could load the books into trays that feed the books to the scanner. It's just a matter of investing in the proper equipment.

[Buzzetta]

Anywhoooo... 

At the most basic, AI should be able to lend a hand to determining page quality with far greater consistency.  As an example I offer the Lego Mindstorms kit that was introduced to retail around a decade ago with the photoreceptor components and sensors which can distinguish color.  One of the first thing the lego community did was create robotic builds mounted with that first camera that could detect the color of and solve, a Rubik's cube.   Videos started appearing using the tech able to distinguish color at mass market retail prices.   Over the last decade Lego has offered upgraded components and software and the community has developed software that continues to push boundaries.  Keep in mind that this is a camera and sensors that are "over the counter retail grade."

I am pretty sure that an upgraded industrial model would be able to accurately detect page quality. 

[Prince Namor]

On 12/24/2023 at 7:48 AM, shadroch said:

Someone correct me if I'm wrong, but I believe the technology already exists to scan a book without opening it.  It would be easy to note missing pieces, handwriting, etc. 

You could load the books into trays that feed the books to the scanner. It's just a matter of investing in the proper equipment.

Exactly. That's exactly how simple this technology is that we're talking about.

Why some people are making it out to be some kind of complex-beyond-human-comprehension thing is beyond me. It simply involves scanning a book and writing a program that looks for certain elements within that scan.

I've seen someone's VAST research in RESUBMISSIONS, and it is just... BAFFLING, how CGC can consistently be INCONSISTENT with their grades. 

[Prince Namor]

Weird to think I'd have to go to Reddit to find some people talking some sense about this subject:

In my opinion, this should already be the case amongst the major grading companies and it's a bit sad that it isn't. Every book should be scanned in high resolution and the image run through a program which notes and comments flaws that it recognizes. A human grader (or two) should then confirm the flaws notated by the AI as well as add any of their own. This would result in more consistent grades imo. You could even have the scan run against an image search and hopefully that would eliminate the mislabeling that has been occurring.

Ya think????

[Prince Namor]

You have to realize that most REAL collectors... the #1 thing that they want... MORE than even increased TAT's is... that CGC will GET IT RIGHT. To be FAIR across the board.

Turn around times are NOT the thing that will eventually... that will eventually HURT this company. It is the INCONSISTENCIES of human error and the FAVORITISM of human GREED...

BOTH that can be extremely LESSENED through the use of AI technology. 

[VintageComics]

On 12/23/2023 at 7:44 PM, CGC Mike said:

I am going to unlock this thread, as I can see where it might be an important subject in the comics industry.  If it goes like the last one I locked, I will be vary generous with points to any and all accounts that I feel are a violation of the terms of use.  Everyone is supposed to be 18 years or older here.  I expect you to act like an adult.  No trolling or any contentious posts will be tolerated.

Thanks Mike. It's much appreciated! 

On 12/23/2023 at 7:45 PM, Chip Cataldo said:

Thanks Mike. That's nice of you to give Roy a Christmas present. 

I talked to Mike about it and he agreed to open it so yes, it's a great Christmas present. 

 

[CGC Mike]

On 12/24/2023 at 12:57 AM, VintageComics said:

On 12/23/2023 at 7:45 PM, Chip Cataldo said:

Thanks Mike. That's nice of you to give Roy a Christmas present. 

I talked to Mike about it and he agreed to open it so yes, it's a great Christmas present. 

To be clear, you asked for the point to be removed from your account and the thread be unlocked.  I told you that the point stays, and I will be reopening the thread.  My decision to reopen the thread did not have anything to do with your request.  It was inevitable that an AI discussion would need to be allowed going forward.  However, I do not take kindly when people open up a discussion, knowing that it had previously been closed a short while back.  The same OP none the less.  

Why am I posting this?   It is due to the fact that I have seen multiple posts of yours saying that you have been working with me and that is why this or that happened.  (not your exact words but means the same)   No forum members are working with me in any way in making decisions on this forum.  People do make requests and I do listen and take them into consideration.  However, I take the time and weigh everything, and make a decision based on my 12 years of experience in moderating this forum.

There is no need to reply publicly or privately, as I will not be replying.  Also, I do not want to see a discussion revolving around this post.  

[VintageComics]

On 12/24/2023 at 8:06 AM, CGC Mike said:

To be clear, you asked for the point to be removed from your account and the thread be unlocked.  I told you that the point stays, and I will be reopening the thread.  My decision to reopen the thread did not have anything to do with your request.  It was inevitable that an AI discussion would need to be allowed going forward.  However, I do not take kindly when people open up a discussion, knowing that it had previously been closed a short while back.  The same OP none the less.  

Why am I posting this?   It is due to the fact that I have seen multiple posts of yours saying that you have been working with me and that is why this or that happened.  (not your exact words but means the same)   No forum members are working with me in any way in making decisions on this forum.  People do make requests and I do listen and take them into consideration.  However, I take the time and weigh everything, and make a decision based on my 12 years of experience in moderating this forum.

There is no need to reply publicly or privately, as I will not be replying.  Also, I do not want to see a discussion revolving around this post.  

As I said via PM, this is a great compromise and I'm very happy with it. Thank you. 

Edited December 24, 2023 by VintageComics

[VintageComics]

As promised:

In regards to a "fair and consistent grading model"

On 12/22/2023 at 5:45 PM, JC25427N said:

That would be great, but I don't think we'll get there even with an AI model, at least not at first. A lot of people think of AI as a soulless and emotionless arbiter that only concerns itself with facts and logic, but that's only true to a certain extent. AI models can still have biases and inaccuracies depending on how they were trained, for all we know once we start rolling out machine learning models for grading people will say its biased towards yellow covers and gives them higher grades because most of the high grade books it was trained on were yellow so it sees yellow as a sign of high grade, or that it's inaccurate at grading Golden Age books because it wasn't trained on enough of them. And that's not even with going into the possibility of someone finding potential exploit in the models (imagine the exact model gets leaked by an "inside man") and people find out how to get restoration past it, how to hide certain defects from it, etc.

Most of these issues could be resolved by using an AI model alongside a human grader, and most of them would probably get resolved as the model matures, but I don't think we're close to a 100% fair, near perfect, consistent, and fast program, I don't think anyones even really working on a first iteration of such a model (You would have to collect a large enough dataset first to even attempt it, and according to some other people's opinions the current capability of hi-res scans wouldn't be sufficient)

 

In regards to not having any bias:

On 12/22/2023 at 9:06 PM, JC25427N said:

This is insanely harder than you make it sound

 

In regards to removing human hands from the equation.

On 12/22/2023 at 9:26 PM, JC25427N said:

Ok, what I'm talking about (and what I thought you were talking about) is a machine learning model (AI) that grades comic books (perfectly, accurately, fast, and without bias as you mentioned before). To give you just ONE of many aspects that make this harder than someone who doesn't work with this kind of stuff may think, you'd have to collect enough data to train this hypothetical model (I'm not even talking about any physical robotics of flipping pages, scanning, turning, etc that VintageComics and everyone else seems to be concerned with, lets just say all of that machinery exists). To even begin the conception of creating a model you need to collect a dataset to train the model on. To do this you would have to either collect a large corpus of graded books (as you said you want this model to be trained only on books that have been seen and graded before), crack them out of their slabs, and then collect all the necessary scans that you would need to grade it, or alternatively. get a large corpus of books, have someone at CGC grade them, and take all the necessary scans before encapsulating them (and this wouldn't just be the front cover and back cover scan that CGC provides now, if you want this model to do full grading end to end you need a scan of every page of every comic book that CGC could ever grade to assure that the book is whole and correct without incorrect married pages, missing inserts, etc). Then you probably need multiple examples of each book at different conditions.  Then you have to know that whatever form of input data is used to train the model will determine what you actually need to give the model when it's used in production for actual grading, so if you give it 50 scans per book for training, you need to collect those scans in production. Collecting this data and collecting it at the quality necessary for good performance (which according to that one person (sorry I forgot the name) who posted earlier who said they have a vast experience in CG and working with scanning technology says that even the most sensitive current scanners wouldn't be sufficient for this task, I don't personally know that, just repeating it). So collecting this data is a herculean task, making sure the data integrity and quality is sufficient is a herculean task, and even storing all of these hi-res scans (again, every page, of every book, at multiple angles, and you might need to repeat this for multiple conditions) would get insanely costly, and I don't even mean just storing the training data, you would probably want to store the scans used in production in case something goes wrong for validation, someone wants to appeal a grade, etc (and then go back and see how many submissions CGC gets).  

And this is just data collection, one of the first steps when deciding to create this model, past that there are other hurdles to jump through. Again just for the machine learning model that goes the grading, I've assumed that the scanning machinery exists (although apparently thats up for debate whether current scanning capabilities are enough, I would imagine it is but I'm not a grader or a CG expert)

 

 

 

In regards to @skybolt 's comment:

I'm just worried about an AI scanning a dust particle that landed on the book and thinking it's a stain.

On 12/22/2023 at 9:28 PM, JC25427N said:

Yeah this is one concern you would have when trying to train this AI, this is part of data integrity, insuring that the data doesn't have any noise (such as dust particles) that mislead the model, creates error, creates bias, etc. 

 

In regards to @jcjames 's comment:

 

AI is currently better, faster, more accurate at scanning medical imaging (which IMO is more complex than creases and spine tics) than any single human.
It didn't happen overnight, but it is possible. 

 

 

On 12/22/2023 at 9:41 PM, JC25427N said:

I think there's a slight confusion here

I think when you say scanning, you mean looking at an image and interpreting it, drawing conclusions from it, etc. Yeah I'm sure AI could do that for comic books if all preconditions are met for training such a model. 

When I say scanning, I mean taking a picture (taking a scan). So the concern raised by some CG expert earlier is that the picture taking technology in the current state is not sensitive enough for training a model to accurately grade. I don't believe it either personally (my main issue in that rant I went on was about data collection with the assumption all necessary machinery exists), but I was just repeating what someone else who claims to be a subject matter expert on CG (which I am not, so I kind of defer to them) said earlier .

 

Regarding using public images and a CGC database of stored images:

On 12/22/2023 at 9:45 PM, JC25427N said:

Alright, I could believe that, but are all those scans of sufficient quality? You wouldn't really know until you train the model and look at the training results I suppose. One thing that could be a source of noise is if you have scans of differing quality in the dataset, that could introduce bias in the model. There's a whole lot of dimensionality to consider here, I'm just going off the cuff with what I would be thinking if I were asked to do this. It would take a lot of work, I'm not saying its impossible, but its not as easy as just saying "train the computer to do this"

 

edit: Sorry I just read all the posts in this thread asking to stop the AI talk, this thread moves at lightning pace forgive me I just saw it

 

On 12/22/2023 at 11:35 PM, JC25427N said:

I work in software development, I just don't think there's an appreciation for how complex it would be to create. I also think some over-estimating might be going on, processing large image datasets can take a really long time, might not be as fast and efficient as people think. 

 

[VintageComics]

Just the topic of scanning and interpreting images, (and this post doesn't even touch on how the book will be manipulated for scanning) is pretty complex. 

The problem as I perceive it it isn't that AI can't just scan and read images, we all know it can, but a book is a 3 dimensional item. 

It's a combination of things like how will the book be delivered and show to the AI program / scanner?

Who will expose both sides of every page in a book (16 - 32+ wraps or 32 to 64+ pages?)

How far open do the pages need to be?

Can you scan them accurately at an angle?

Can you scan them accurately with curvature on the pages (meaning they aren't laying flat)?

How do you inspect staples? Especially in square bounds?

You're already imaging 32-64 pages, how do you image the sides of the book, like the open edge, the spine the tops and bottoms?

I've had books with defects ONLY apparent on the open edge of the book - literally across the face of the page ends. Really. I had a book where someone wrote in ink that only showed up in UV on the literal edges of the pages and the defect wasn't visible from the flat side of the page. 

This was one of those strange defects that I was talking about that you would have to completely retrain AI for once you uncovered a new flaw. 

Then we have the problems of stuff like specks of dust and other things that can corrupt images as discussed above. 

Every little new factoid is a learning curve for AI that would take countless hours to adapt to, that humans can factor in, in a matter of seconds through a personal discussion around a grading table. 

 

These are just some of the things that I've written "off the cuff" so to speak. There are probably a zillion others. 

 

[Buzzetta]

Remember that when it comes to third party grading that baseball cards have always been on the forefront of the technology, presumably due to lack of complexity in grading a card compared to other collectables.  AI is making an appearance in card grading.   As the technology progresses, it is only a matter of time before some aspects will find their way into grading other collectables. 

https://agscard.com/

Now whether or not it is successful remains to be seen.  However the tech has been implemented. 

[VintageComics]

On 12/24/2023 at 11:09 AM, Buzzetta said:

Remember that when it comes to third party grading that baseball cards have always been on the forefront of the technology, presumably due to lack of complexity in grading a card compared to other collectables.  AI is making an appearance in card grading.   As the technology progresses, it is only a matter of time before some aspects will find their way into grading other collectables. 

https://agscard.com/

Now whether or not it is successful remains to be seen.  However the tech has been implemented. 

It's been implemented in cards for 2 specific reasons that I can think of.

1) cards are much easier to grade than books

2) the money in cards dwarfs the money in comic books. Remember when $1 million comic book was crazy money? There were already multi-million $ cards floating around. 

Again, I'm not saying AI will never grade comics, I'm just saying making it accurate, cost efficient and time efficient are feasible over all, are the complicating factors.