Kouji Mori reveals they considered using AI to draft new Berserk episodes

[scaler26]

In a new interview, it is stated that AI was considered in the production of the new episodes:

https://news.yahoo.co.jp/articles/ae91ca48bb4da9147789a6a462367720db05ba56?page=3

(Yahoo! Japan is no longer accessible in Europe, so any fellow Europeans that want to take a look will unfortunately need to use a VPN)

I'm using (ironically enough) a machine translation, which might not be perfectly accurate, but the general gist of it is that they'd input Miura-sensei's art and use generative AI to create drafts for the continuation. They eventually ditched the idea because they were "unable to find a way forward for [this] project, and decided to proceed with the conventional drawing method."

Now, the fact that the thought crossed their minds is bad enough, but that using derivative drivel made by some program and not a human being was thought of as a viable option in "honoring Miura's memory" is mind boggling, and yet another example of this team's creative bankruptcy. Moreover, the decision to not use AI was seemingly not born out of any moral principle, but their dissatisfaction with the results. The question that now remains is: Given how fast this technology has been evolving, and that they intend to stretch the continuation for the next decade, will there be a point when AI is "good enough" for them? Is this what we have to look forward to now?

I think this is all pretty damning, and if there were any doubts about the sincerity of this project, surely they will be quelled after this.
I had already decided I wouldn't be buying any volume released after Miura-sensei unfortunately passed away, but this is the cherry on top.

 

[Aazealh]

Mori's numerous interviews about Berserk tend to be pretty self-serving in my opinion, but I wouldn't pin this specific thing on him. This sounds more like something Kurosaki would attempt. I also think it underlines their thought process for this whole endeavor: trying to "keep Miura in" the continuation, even through wrongheaded means. To me, it's another indication that Mori's original idea (doing a written summary with illustrations) was the right one. Simpler and purer.

Anyway, other than this tibdit, I found a number of interesting new details in the slew of interviews Mori did for the release of volume 42.

 

[scaler26]

I wouldn't pin this specific thing on him

This was a good clarification, I definitely could have worded it better, my intention was only to say where this piece of information came from, not that I'm blaming him

Thank you for taking the time to make this summary of pertinent info from all the interviews!! I hadn't seen the Da Vinci interview.

 

[Aazealh]

This was a good clarification, I definitely could have worded it better, my intention was only to say where this piece of information came from, not that I'm blaming him

Don't worry, I get it. It's just my own personal guess anyway, and I'd say it's fair to attribute responsibility to all parties associated with the project.

Thank you for taking the time to make this summary of pertinent info from all the interviews!! I hadn't seen the Da Vinci interview.

No problem. Walter suggested I move it to another thread, so I've left a link to redirect instead.

 

[blodvind]

but that using derivative drivel made by some program and not a human being was thought of as a viable option in "honoring Miura's memory" is mind boggling

I think that would be a better option, actually. If not right now, in the near future. It is possible that an AI will get the style closer to the original than what the team is doing.

The value of human- vs. machine-derivate material is another philosophical question, and we should start considering that an AI can be indeed creative. This technology is going to humble us humans down very quickly, I think.

 

[Walter]

I think that would be a better option, actually. If not right now, in the near future. It is possible that an AI will get the style closer to the original than what the team is doing.

The value of human- vs. machine-derivate material is another philosophical question, and we should start considering that an AI can be indeed creative. This technology is going to humble us humans down very quickly, I think.

Even setting aside the extremely inconsistent capabilities of the technology, which are nascent to say the least, it would have been ethically fraught to continue Berserk with AI.

That being said, it is not damning in my opinion that they considered but ultimately did not pursue AI. It was merely an (bad) idea brought up for discussion it seems. Whatever. Add it to the pile of bad ideas that were involved.

 

[Aazealh]

The value of human- vs. machine-derivate material is another philosophical question, and we should start considering that an AI can be indeed creative. This technology is going to humble us humans down very quickly, I think.

You don't seem to know how these things work. GenAI models aren't creative, they are computer programs trained on large quantities of artwork which they can then recombine in a variety of ways following user prompts. Presenting them as alternatives to human art is fallacious since they cannot exist without it.

More importantly, the kind of craftsmanship involved in creating a manga like Berserk goes way beyond approximating artwork, and is completely out of reach for this technology. In fact, it's this very lack of craftsmanship that's problematic with the continuation in the first place. Lastly, do note they were talking about using it for drafts, not the actual finished drawings.

 

[AdvancedStupid]

The real artist is dead, we all know this. No one is going to fill those shoes. So why lie to ourselves?

Japans corporate/business structures is one of the most ossified in the world. I honestly think it was more hype than anything to mention that they "looked into AI" because if they actually sat down and did look at it they would have concluded that they can offer a much better product by shifting the budget around in this manner. Now things release in like a half-year cycle and is nothing more than a bastardization on all fronts (which from their perspective is business as usual so they win we lose I guess?)

They may not be able to craft the story in any unique or cleaver manner using an LLM but they can speed up art production (free up money that would have gone to artists to hire WRITERS instead). I mean just take a look at some of these pictures generated without even mentioning Berserk (can't mention copyright IP to the free version of these things). I first asked ChatGPT to describe to me in verbose detail how the art style of Berserk is expressed, than I copy-pasted that into Dall-E which created the following (with specific PERSISTENT training you can mimic the exact style, this is zero training, just morning posting before going to work):

Spoiler: LLM generated picture, you have been warned

I'm not an "AI" developer (I keep it strictly C++ and native programs; running an application through a browser tab is the antithesis of everything I stand for) but I think the studio should have looked deeper into this technique.

Make a single, but much larger (200+ pages), Berserk manga release to conclude the story by leveraging these generative models to do the heavy lifting in the artwork allowing the company to shift the budget to hire more writers to focus the story, dialogue, and even edit the panels appropriately.

They can hire a contracted community manager to reach out to places like this to get a pulse of what people expect from the manga. It's all budget in the end.

Regardless you'll need 1-2 artists to correct the mistakes "AI" generates sometimes (it seems much less when it comes to non-photo realistic art but still there) and to do very specific pieces of art. Not to mention sitting on top of the application, highlighting parts that need to be regenerated and then editing what needs editing (digital composition).

Heck if they wanted to go even further they can do what students are doing today by training the data not on the art but the private notes & journals of an individual. Train a model on Miura's notes, journals, interviews, etc and that model can answer questions such as whether or not the studio is staying faithful to the vision of Berserk. Though I don't know if Miura kept any such things to qualify, most students I see use it on academics who kept libraries of private notes, essays and journals, some approaching a million words (ie https://chatgpt.com/g/g-6ihgGYf1R-hg-wells-chatbot H.G. Wells, 1.3 million words or https://chatgpt.com/g/g-MDDd0UedM-ludwig-von-mises-chatbot for Ludwig von Mises, 800k words).

It's just a shame, discovering Berserk 3 weeks ago, reading all the mangas to only hit this point and find out it has become a husk of a zombie shuffled away by Japanese-style corporate decision making (no one ever asks why Japan doesn't have a tech sector equivalent to China or the US even though they were a technological powerhouse in the 80s; because when you have an ossified management structure it ripples down hard).

In time these free LLMs will be enough to allow end-users to just make their own fan fiction and be done with it. One cannot stop progress. There is a dull pain in knowing that there never will be a real canon ending.

 

[Aazealh]

The real artist is dead, we all know this. No one is going to fill those shoes. So why lie to ourselves?

You do seem to be lying to yourself here, I'm afraid, or at least you've misidentified where the difficulty lies in continuing Berserk without Kentarou Miura.

Your premise is that it's just a budget issue: if they could hire more writers, and even a community manager to ask fans what they think, then it'd be no problem. But... That's really, really not how it works. Hakusensha already has a bunch of people deciding on the story (Mori, Kurosaki, people from Young Animal...), and they could hire more people if they wanted to. They have the money.

They don't do it because it wouldn't help. Since you're a developer, you must know about the "mythical man-month", and specifically the specious notion that if one person needs 10 weeks to do a task, then with 10 people it'll take a single week. That doesn't work for software development and it works even less for creative endeavors. For the same reason, trying to shape the story based on Internet feedback would be catastrophic.

If Berserk was great with only one guy at the helm, it's because said guy was great himself. No amount of fancy tools or cooks in the kitchen will emulate him.

It's just a shame, discovering Berserk 3 weeks ago, reading all the mangas to only hit this point and find out it has become a husk of a zombie shuffled away by Japanese-style corporate decision making (no one ever asks why Japan doesn't have a tech sector equivalent to China or the US even though they were a technological powerhouse in the 80s; because when you have an ossified management structure it ripples down hard).

In time these free LLMs will be enough to allow end-users to just make their own fan fiction and be done with it. One cannot stop progress. There is a dull pain in knowing that there never will be a real canon ending.

I get that you're disappointed by the current situation, it's a shared sentiment believe me, but there's no silver bullet for it. No magic button one can just press to "fix" it. Creating manga is complex and difficult in and of itself (it's not a matter of putting illustrations in panels), and producing a work of Berserk's quality is nigh impossible. Expecting that someday anyone will be able to do it just by asking their computer is a self-delusion.

In addition, I don't want to talk at length about the limits of large language models or diffusion models because I think it's besides the point, but you should understand that there's a hard cap to what you can get them to do. While OpenAI and others are hyping them up (because that gets them funding), they don't have limitless potential for improvement.

 

[Xenoskell]

I mean just take a look at some of these pictures generated without even mentioning Berserk (can't mention copyright IP to the free version of these things). I first asked ChatGPT to describe to me in verbose detail how the art style of Berserk is expressed, than I copy-pasted that into Dall-E which created the following (with specific PERSISTENT training you can mimic the exact style, this is zero training, just morning posting before going to work):

Spoiler: LLM generated picture, you have been warned

Tons of design errors in these illustrations.
The elements aren't separated by layers (such as one layer for background, other for screentones, armor/clothing, etc). There's also no context, the images doesn't seem to have any connection with each other.

It would take much more time to correct everything than just making it from scratch.