Krita Commitment to Empowering Human Artists (CEHA)

[Tymon Dąbrowski]

I'll try to make it short. [Edit later: I'm sorry. I really tried.]

*1) Artists *(those that protest)* are against generative AI.* I don't
think we should conflate generative AI with all possible AI - for example I
doubt they are against all those new safety features in cars. I've seen a
car parking itself in real life, that was amazing. Likewise, I do believe
there are plenty of possible usecases of AI that won't have the negative
traits (steals copyrighted work, steals jobs from artists etc.) that
generated AI is criticised for.

*2) What projects are suggested for Krita.*

There are two suggestions.
2a)* Automatic tweening* for animation. Suggested by Dmitry. From the user
point of view, they'd have some animation, wanted to have more frames per
second, and AI would generate the frames in between. Links:
https://github.com/google-research/frame-interpolation ,
https://arxiv.org/pdf/2202.04901.pdf

2b) *Automatic inking/lineart*. Suggested by me. From the user point of
view, they'd have a pencil-like sketch, and AI would generate lineart. Of
course it wouldn't fix all the little issues artists are often fixing
during inking, but it should greatly shorten the time of the repeatable
work. Link: https://esslab.jp/~ess/publications/SimoSerraSIGGRAPH2016.pdf
Note that the original article only creates fineliner lineart from pencil
sketches. I am kind of thinking of making it more universal, allowing for
more styles.

I'm a bit less sure about 2a) as I didn't really research it thoroughly
yet. On the 2b) project, I believe we can do it in a way that should be
considered fully ethical: (i) ethical dataset*, (ii) won't be stealing
jobs, most probably. Ad (i), the original dataset had 68 images; I think we
should be able to gather a similar amount from either donations
(preferable) or CC-0 images. Ad (ii), I know that comic books often have an
"inker"; however many times they work traditionally, and I don't think our
AI would replace even the digital ones (who often add a personal, distinct
touch to the lineart). I do think that it would mostly help the artists
that ink their own projects and whose lineart is just a part of their
style, not a statement piece (for example in the modern anime-illustration
style lineart is often very much in the background) (To see what I mean,
please compare this Batman illustration:
https://www.deviantart.com/kriss777/art/Batman-Sketch-by-edbenes-Inked-252674704#comments
to this anime one: https://www.artstation.com/artwork/48Z0PW ). I think the
effect of AI's work would just be the first stage of inking, and would only
replace the boring, uncreative part of the work.

I don't think our network would be even capable of steering too much from
the sketch, let alone be creative in any way*, so I doubt it would replace
actual artists. (* I just think the network and dataset would be way too
small for that. I'm less sure about the 2a project because I don't know
that much about it)

*3) What I think we **as Krita** should do regarding AI.*

Generative AI is a new scary thing (it will change things, but we don't
know how; unknown is scary, sometimes for very good reasons, sometimes
not). NFT was dumb, useless and terrible for the environment - generative
AI isn't as dumb or as useless. I think we should just wait and see how the
world looks like in a little while.
I think we shouldn't plan on adding generative AI for now.
I think we can say that we aren't planning on that.
I think we can make projects that utilize AI in a safer, more predictable,
smaller way. When doing so, we need to be *very very clear* that we are
*not* using copyrighted images and that we are *not* planning on adding
generative AI for now.
"We're not planning on adding generative AI any time soon" is good because
it's very clear and precise and true. I feel like Emmet's MR was a bit too
broad and vague.


*4) Answering questions.*

--- Ethics - answered above.
--- Usefulness - I don't know. I think we should ask.
--- Practicality - for effort estimation, I think maybe Dmitry would answer
better. I know that some people in uni did projects like style transfer in
one semester and it looked very decent, and it was just one project, not a
thesis or anything. But we need to do all the GUI etc. too, so it might
take longer. Dependencies - that's a good question. Dmitry said that
OpenVino should cover all CPUs - in such case, that would be the only one
required for deployment (the whole training part doesn't have to be
included in Krita). I hope it would be fast enough that users won't need to
ask for GPU support - that's something we'll know from initial experiments
(it probably would have to be included for the frame
interpolation/inbetweening project though, since there would be more
pictures to process?).
--- GPLv3 - I don't think that a model is "not changeable". First, users
could probably very easily just use the article's authors' model instead of
ours. Second, they can train their own. Third, they can take ours, and
retrain it (unlike the article's authors' model, which is non-commercial,
we can make ours just fully GPL licensed, afaik). Considering that (at
least in case of the inking project) our model will be fully convolutional,
frankly any fully-convolutional model should work without many changes,
since it won't have any special requirements for input image sizes. Fourth,
with some work it would probably be doable to just change a specific number
in the model. True, you might need a bit more expert knowledge to operate
on the model, but we provide Krita's surce code without expecting most of
our users to know programming, too. I might be misunderstanding the
question.
--- Reputation & Future Proofness - the question of reputation is very
important, yes. I don't want to hurt Krita's. Respectfully though, unlike
NFTs, I don't think the majority of users of generative AI are "tech bros".
Generative AI is custom-made beautiful art, if you ignore all the context.
It can show your dreams on the canvas in the real world, and now you can
show them to your friends and family. People love art, and most people
can't make it themselves, especially not to the quality they'd want to see,
and often won't afford or justify the price of a custom-made artwork.
People have vivid imagination and generative AI is one of the ways to make
it real. If you think about it in isolation (pretending there are no
ethical problems with it etc.), there are plenty of reasons why an average
person (or an artist) would want to use it. Writers, to make cover art for
their books. Game developers, to have pretty art in their games. DnD
players, to make their gameplay more interesting. Office workers, to make
their presentation less generic and boring. Architects, to see the basic
idea quickly, without building it painstakingly in a CAD program or
Blender. Artists, to see a draft of a project before putting much work in
it, or to finally be able to find the perfect pose reference. Midjourney
apparently has ~16mln users on Discord. Those are not just tech bros. I'm
not defending it, I'm just saying, thinking about them as "just tech bros"
is simply incorrect.
About how many professional artists use generative AI, I have no idea.
Still, there are other types of AI too.





pt., 16 lut 2024 o 15:40 Halla Rempt <halla at valdyas.org> napisał(a):

> I found it quite well-written, actually, and a good starting point for
> thinking about AI. I especially like the focus on making clear that Krita
> is artists, not for art.
>
> But before we do anything with AI we need to consider a number of points:
>
> * Ethics. Obviously using models based on stolen data is a no-go. Adding
> AI slime that is already drowning the world is also a no-go: Krita should
> always focus on helping the artist do their work, not replace the artist.
>
> * Usefulness. Do artists really want the features we're considering?
>
> * Practicability. Is the effort needed to integrate i.e. automatic inking
> in Krita worth it? How much effort would it really be? What about extra
> dependencies.
>
> * GPLv3. Is the inclusion of a blob that the user cannot replace, i.e. a
> model, not against our license. Krita's functionality needs to be
> changeable by the user, after all.
>
> * Reputation. Like it or not, but the people who like AI most are not
> artists, but tech people, or people who use images casually next to their
> real work. Will there be a backlash if we spend time and money on AI?
>
> * Future proofness. Is AI a fad, like NFT's were, or something with a
> future?
>
>
> On vrijdag 16 februari 2024 05:36:45 CET Emmet O'Neill wrote:
> > Hey all.
> >
> > In light of this week's discussion about AI/ML in Krita, I want to share
> > this meta MR about adding a statement of intent to the project's source
> > tree which outlines a few basic commitments that I believe we should make
> > in the interest of benefiting our users and fostering a culture of human
> > art.
> >
> > I've outlined what it is, what it isn't, and why I think we should commit
> > to it in the body of the MR.
> >
> > https://invent.kde.org/graphics/krita/-/merge_requests/2071
> >
> > Emmet
> >
>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/kimageshop/attachments/20240218/6876824b/attachment.htm>