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

[VintageComics]

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

Have at it! 

[VintageComics]

I personally have stated many times that AI can do aspects of grading, but grading an entire comic is not going to happen anytime soon for a zillion different reasons, but the primary one being that you need to intersect the following

i) handling the actual comic book is complex

ii) time constraints - sure, if I had an hour to feed all my info into a computer I may be able to get a reasonably accurate grade, or grade range but you need this done in seconds

iii) cost constraints - how do you build something this complex, in a space that doesn't take too much room, is cost effective and accurate

iv) you will ALWAYS need some level of human interaction throughout the grading process, whether it's from receiving packages, to unpackaging, to laying them down on an assembly line, to removing them from Mylars, to grading them (flipping through all pages) to putting them back into Mylars, etc. so where does the human stop and the AI start and visa versa?

v) you need to TEACH the AI to learn along the way (or from past experience) - that's a stall every time you find a new problem to solve, and this happens daily when grading 1000's of books a day. For example, there are things that happen in the grading process EVERY DAY when weird defects, or a crossover of defects happen that need actual human conversation to make a decision. This will stall the grading process and needs to be learned and it happens daily.

vi) you'd need to make AI changes every time a new defect or variant is discovered - new learning curve

----------------------------------------

The way I see it the variables are myriad and not easy to solve. 

Humans do this sort of learning very quickly. 

I just don't see AI grading books accurately, in a cost effective and timely manner anytime soon. 

Thoughts?

[Prince Namor]

Nobody is talking about robots that handle comic books. It's a scanning process that thinks. It takes away the most biased aspect of grading from a human.

[MattTheDuck]

On 12/22/2023 at 6:45 PM, VintageComics said:

you need this done in seconds

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.

[Prince Namor]

On 12/23/2023 at 10:12 AM, 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.

Exactly. You'd also eliminate the possibility of someone knowing someone at CGC and being able to get 'favors' done.

Faster, more efficient, no bias, never gets tired... 

[lizards2]

quit posting this shatnizzle, and sell some nice raw books!  

[JC25427N]

nevermind, I'm not gonna try anymore, I don't think I'll get anywhere. Good luck

Edited December 23, 2023 by JC25427N

[Axelrod]

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

I personally have stated many times that AI can do aspects of grading, but grading an entire comic is not going to happen anytime soon

Thoughts?

I mean, what is "soon?"  

It's kind of reductive to say that they currently don't have the tech for ("pure") AI grading now. That's obviously true.  

But I think it's impossible to claim that AI would never be able to do it.  It can definitely scan the front and back covers, and that's like 90% of the grade.  Could also give a far more objective determination of page color than what happens now. 

Currently you need a person to go through the book and check for interior defects.  I will allow that's much more complex to 100% automate, but I would never say it couldn't be done either.  

[LudaToke]

You don't even need to implement AI to pull off something like this.

I used to work for a Pavement Management company that had vans that were outfitted with cameras, lasers and scanners that would collect images of the interstate while travelling at the speed of traffic (55 mph).  The images were then run through software that would produce readouts that would shows cracks and pavement defects that would allow for the various states to determine which areas were defective and needed to be replaced. We were able to collect and analyze hundreds of miles of data in a single day.  A single comic book analysis would be minute compared to this. This was well over a decade ago and the company had already been doing it a decade prior to that.  I can imagine the same image capture and analysis technology has advanced quite significantly since them and didn't involve AI at all.

Sometimes you don't need to kill a fly with a bazooka and the solution is much less involved than you think.  The cost of one of these vans a decade ago was about $250K so it wasn't terribly expensive either.

[VintageComics]

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

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

you need this done in seconds

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.

So you want to "build a million machines" that don't exist yet? This sounds economically feasible?

Or are you going to have 1000s of graders flipping pages for these photo-imaging machines so that the books can be scanned and graded?

So we'll pay people to stand around and just flip pages? Or load them into robotic machines that flip pages?

 

I can't believe some of the ducking (pun intended) ideas some people come up with without thinking them through.

I guess we can do it this way:

[VintageComics]

On 12/22/2023 at 11:55 PM, JC25427N said:

nevermind, I'm not gonna try anymore, I don't think I'll get anywhere. Good luck

I've got you. 

Too late now, but I'll copy your posts from the ASM #252 thread into here tomorrow. You have made some incredibly salient points about AI and are pretty much the only reasonable voice in the entire discussion. Thank you. 

Edited December 23, 2023 by VintageComics

[MattTheDuck]

On 12/22/2023 at 10:42 PM, VintageComics said:

So you want to "build a million machines" that don't exist yet?

Well, nothing like that has ever happened before, so perhaps it's a pipe dream.  Maybe someone will come up with a way to inexpensively mass produce things.

[VintageComics]

On 12/23/2023 at 1:47 AM, MattTheDuck said:

Well, nothing like that has ever happened before, so perhaps it's a pipe dream.  Maybe someone will come up with a way to inexpensively mass produce things.

This isn't NASA. It has to be feasible for a business.

That means cost, efficiency, learning, adaptability, size, scalability, automation, productivity, profitability, all of these things and more need to be considered. 

We have @JC25427N trying to explain JUST the software aspect and nobody seems to understand the complexity, and there are probably dozens of other things that need to be considered and co-incide. 

I don't understand the software but I DO understand the complexity. 

You'd basically pretty much need to recreate a human to do a 3 dimensional task. It would need to have very fine motor skills, or (again) you're paying a human to do the motor work, slowing down the process and adding human error, thereby in ways (again) corrupting the process.

After discussing this for several hours I am fairly certain most people have no clue how complex of a solution that's being proposed this is. 

I've already been arguing for years in the minority that it ain't happening and my mind still hasn't changed. Nobody yet has been able to produce a reasonable solution. 

Edited December 23, 2023 by VintageComics

[MattTheDuck]

On 12/22/2023 at 10:56 PM, VintageComics said:

This isn't NASA. It has to be feasible for a business.

That means cost, efficiency, learning, adaptability, size, scalability, automation, productivity, profitability, all of these things and more need to be considered. 

We have @JC25427N trying to explain JUST the software aspect and nobody seems to understand the complexity, and there are probably dozens of other things that need to be considered and co-incide. 

You'd basically pretty much need to recreate a human to do a 3 dimensional task. It would need to have very fine motor skills, or (again) you're paying a human to do the motor work, slowing down the process and adding human error, thereby in ways (again) corrupting the process.

After discussing this for several hours I am fairly certain most people have no clue how complex of a solution that's being proposed this is. 

I know everyone thinks I'm dumb, but I've already been arguing for years in the minority that it ain't happening and my mind still hasn't changed. Nobody yet has been able to produce a reasonable solution. 

Yep, it's impossible.  Completely agree.

[Prince Namor]

No one is making a case for Futuristic Robots to do the entire process. Complete straw man argument.

An employee scans the book into a computer program. The computer does the basic grading. It's that simple.

The employee counts pages, and examines to make sure nothing is missing inside.

A grade is assigned.

Major League baseball is looking to test an automated strike zone in 2024. They have the technology to trace a 100 mph fastball into a three-dimensional INVISIBLE box that measures 17 inches by about 3.5 ft. 

Already AI can detect brush strokes in classic paintings and detect forgeries and the differences in style.

Know why each of those are interested in that process? Because BILLIONS of dollars are at stake.

[RockMyAmadeus]

[s-man]

At first I thought who is this Al person, and why not have him grade comics?  And then I was like, oohhhhhhh A.I., like the movie 

[Get Marwood & I]

On 12/23/2023 at 2:35 AM, VintageComics said:

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

Have at it! 

Haven't we had this discussion already Roy - in the CGC QC thread and also in a separate AI thread that you started?

Based on the CGC article below, I suspect the technology exists to allow them to safely grade moderns without much / any human intervention. Page count could be assessed by weight, and the software can detect sizing and cover flaws. In the cases where CGC slab moderns, fresh from the printers, what percentage of them do you think would have internal issues like missing pages or coupons? If historic data, say, showed that less than 0.001% of moderns have page count issues, would any business still count them? If a modern is missing pages, it will flag by weight in the same way that supermarket self-checkouts do. Or it will get slabbed, and no one will ever know unless they crack it out. Who cracks open moderns, having purchased them in a slab, or sent them in to be slabbed?

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.

Think about it - how exactly did Carney's AI platform eliminate a backlog of 1M collectibles, if not by grading them too? If it identified 1M collectibles in a flash and sent them to work queues, people would still have to manually grade them so the backlog would remain. Or do we believe that CGC had graders sitting around waiting for collectibles to be assigned to them? 

People are always commenting on how their 9.6 should or could have been a 9.8, and vice versa. We're all open to the possibility of two different graders arriving at two different grades. So for moderns, would you build a manually intensive grading operation if you knew that AI could do the majority of the work and hit the same error margins as people do? People will likely always grade the older, expensive books. What company is going to pay people to check that the latest Spider-Gwen has the right number of pages?

https://www.cgccomics.com/news/article/11945/

"Carney has been the instrumental leader in building a proprietary platform that enables CCG to identify a collectible from among millions of possibilities in a near-instant using artificial intelligence (AI), machine learning and advanced computer vision and image processing technologies."

[mjoeyoung]

Exactly.  Nobody here is arguing that AI is going to be dropped in tomorrow and start grading on an Action Comics #1.  You don't need to design a system for the exceptions, you automate the easy volume work, like the moderns, and then, as technology improves you expand the population that can be graded in this way.

On 12/22/2023 at 7:45 PM, VintageComics said:

v) you need to TEACH the AI to learn along the way (or from past experience) - that's a stall every time you find a new problem to solve, and this happens daily when grading 1000's of books a day. For example, there are things that happen in the grading process EVERY DAY when weird defects, or a crossover of defects happen that need actual human conversation to make a decision. This will stall the grading process and needs to be learned and it happens daily.

 

I think this is the biggest flaw of the CURRENT system, too much ART and not enough SCIENCE.  A defect is a defect.  Quantify it and move on.  Let the buyer make the Qualitative assessment.

[Get Marwood & I]

On 12/23/2023 at 11:30 AM, mjoeyoung said:

A defect is a defect.  Quantify it and move on.  Let the buyer make the Qualitative assessment.

Exactly.

AI can grade books like this one below and no one is going to complain about the resulting grade any more than they did when humans were doing it. 9 point something will do. The chances of there being something nasty inside - and it then being discovered - is almost certainly low enough to accept the risk on.

As for older books, the grading contests show that a group of people can more or less get within a grade or two of a final CGC grade by looking just at front and back cover scans. And those that miss by two increments say that CGC got it wrong, not them. We all know roughly what a 4.0 looks like, a 6.0 an 8.0 and a 2.0. Give AI enough examples and who knows, maybe it could then make assessments that would be near enough to go widely unchallenged. If it is a grade or two out, who is going to put that down to AI assessment failings? Humans are a grade or two out, day in, day out, and we all live with that.

So, does you book matter? That is to say, is it worth anything?: Human grading intervention likely.

Does it not?: AI grading, with built in content risk acceptance likely (which is no different to QC not being undertaken. We pay to have it inspected, and we see examples all the time where it clearly wasn't).

Anyway, that's the plot of my new Marvel film sorted. "I, Robot Comic Grader". All I need now is my Diversity Check List satisfied and I'm off and running to losing hundreds of millions of dollar pounds