Experiment: Stable Diffusion Ai (left) and Krita paint-over (right).
In short, this process is tedious. Fixing anatomy issues of the Ai takes a lot of time. Also, some parts are just too random to make sens (she eats chocolate? ) but I'll keep testing, many new possibilities!🍫

@davidrevoy Love to see people taking AI art as chance, not threat! ❤️

@Natanox 🙂 Oh yes, it's a big chance. I can't wait to storytell more and paint less tedious part of art that can be automated. (part in backgrounds, or anything a super Ai assistant could help). I already absorbed in my workflow digital-painting when it was fresh and new, 3D model paint-over, photobashing paint-over, and I'll reserve the same treatment for Ai assisted painting. I understand the one who are against; probably a mind set that put all innovations as a risk.

@davidrevoy @Natanox My worries with AI in the art space are mainly that so many tools are incredibly proprietary but are marketed as open. Not to mention that we mostly don’t know how much of the data that went into a dataset was taken without consent. That stable diffusion is seemingly openly available is a good trend, though, and using it as assistance or inspiration seems to have lots of potential.

@peacememories @Natanox

Yes, the dataset is an issue, and generated art can be also infriging copyright holders (eg. no idea what Disney might think about ) . Artist are suspicious that the dataset already use large part of Artstation and DeviantArt. For ArtStation; Epic Games boot it not long ago, and of course have a license on all the art here. That could make them in position of offering a commercial tool that use 'Ai' with their database.

@davidrevoy @Natanox another story of big companies using free labor to consolidate their position of power… i do hope we will get a healthy space of open datasets and constructive use of machine learning in art.

@peacememories @davidrevoy Actually right now it gets worse. The EU is trying to create ruling for AI and right now awfully fails, as the safety of AI tools shall lie on the *creator* instead of on those who implement it. This means open-source AI projects can be freely used by companies, but when it outputs something troublesome the maintainer receives the lawsuit, not the company, leading to - again - a monopoly on tech.

I'm so fed up with the world we live in right now. Greed always wins..

@davidrevoy I see. Given backgrounds are / can be more abstract anyway it's really handy indeed. Especially with Stable Diffusion where you can give it a really rough version of a landscape to add all the details etc.

I guess people are simply scared as always, like when books, radio, movies, 5G etc. came around.
I just wish they were as sceptical in regards to what big tech companies is doing to them. AI is not the problem, the way it's may be...

@davidrevoy Ça donne presque l'impression d'un GitHub Copilot ou OpenAI Codex, mais pour la peinture numérique... 🤔

@helene Haha, oui, le parralel est bien trouvé. Ceci dit, c'est ci pété que ça le code généré par Github Copilot? 🤣

@davidrevoy De mon expérience brève, assez souvent oui... Mais parfois ça a des éclairs de génie, sur certaines choses !
@helene @davidrevoy En même temps, vu que même le code généré par un programme simple (genre grub-mkconfig) c'est pas mal pète gueule, alors quand c'est du machine learning…
@lanodan @davidrevoy C'est quand même plus propre en sortie de machine learning; ça ressemble beaucoup à du GPT-3 sauf que c'est du code et que ça compile et reste globalement cohérent.

@helene @lanodan Ha je vois. Là, ce que je trouve moyen c'est que les database des Ai images sont clairement entrainé via des mega base de donnée aspiré de ce qui semble provenir de Artstation et DeviantArt. Ca pose le même problème de license éthiquement que ce qui se passe sur Github avec le code.

@davidrevoy @helene Ce qui est un peu dommage, de l'art graphique sous domaine publique c'est pas trop ça qui manque, même si c'est pas très moderne (genre pas de style animé post-2000 avec des couleurs très vives).
@davidrevoy @helene @lanodan J'ai noté qu'en ajoutant "vector art" aux descriptions que je donnais à un générateur d'image utilisant l'algorithme stable diffusion, il me reproduisait le watermark d'un site d'images stock sur l'image générée.

Outre le problème de blanchiment de licence, un autre problème que j'ai vu soulevé est celui d'une confusion sur l'attribution des œuvres, qui s'apparente aux contrefaçons. En effet, utiliser le nom d'un auteur dans la description pour copier le style lors de la génération d'image, puis publier l'image en laissant la description utilisée lors de la génération peut provoquer une confusion en faisant croire l'image a été produite par l'auteur indiqué.


En effet, un simple regard sur démontre un géant soucis.

L' Ai c'est manifestement entrainé sur des oeuvres non libre. Aussi, cette galerie de nouveaux moebius est attribué à son nom semantiquement dans les 'prompts'.

On peut imaginer les moteur de recherches très vite incapable de dicerner un vrai Moebius d'un "Moebius d'Ai". Et aussi un blanchiment; quel est la license des oeuvres générés? Pas celle des ayant droit de Moebius...

@helene @lanodan

@davidrevoy @Feufochmar @helene
Wow, super example.

Et oui, license, automatisation ou non, la respecter devrait être obligatoire, surtout pour ce genre de cas où le "fair use" passerais sans doute pas (et de souvenir faut au minimum une attribution).
Et ça serais pas une surprise d'avoir déjà eu des cas du genre dans le passé avec des outils comme Google Translate pour brouiller les pistes.
@davidrevoy thanks for the insight. Nice to see exploration of this concept.

@davidrevoy Wait, what? You're using Adobe Illustrator?! 🙀


@davidrevoy ah well, non-square images still occasionally mess up body parts. At least faces are relatively consistent now with Stable Diffusion. Disco Diffusion used to make a complete mess of faces.
Its also a good idea to just generate a bunch of images and then cherrypick the best. SD is so fast that it's not really an issue anymore.

@sml Hi, thanks! Yes, maybe it can speed up for hands and specific background objects. I'll check the Inpainting workflow.
For faces; I think I'll always want to tweak them until they fit my style (or until Ai starts to be trained for my style).
All in all, I think paint-over to obtain a final piece is difficult, but good for quick setup and discovering ideas.

@davidrevoy but there isn't any anatomy issues because it's exactly how the #Ai sees humans...😅😉

@davidrevoy Have you tried adding your name or training on your own work (if SD is not already trained on P&C)? This might be an obvious answer as I am on mobile.

@thomzane No, I haven't, I just went for witch on the general Lexica catalog to see if Pepper had an influence on the database. I can't spot anything.
I wanted to install stable diffusion, but to get a local database for the training, the website required a login and I put that to pause at this moment.

@davidrevoy Very interesting! I'd be interested to know how much time was taken out of the planning stage for this to find something that looked good enough to paint or come up with inspiration. Especially if you'd never think to even paint something like this.

@jookia That's where Lexica makes it easier for artist; and also less CPU intensive (when many prompts can be already catalogued)

@davidrevoy I'm very excited to know you're looking at how to integrate AI art in to your process or how it won't integrate.

But I'm also kinda annoyed that I spent like an hour doing Pepper AI requests begging for it to make a witch hat properly to which it said 'no'. Writing prompts to get an AI to output something is definitely a time sink and skill that needs to be factored if we factor in creating the AI art.

@jookia Yes, I played with the webUI , engineering good prompt sounds sometime to be an activity that could take me more time than starting to paint the thing 🤣

@Baa @davidrevoy wow the new episode of pepper&carrot is taking a VERY dark turn

@davidrevoy haha the more I look at the first one the more uncanny / confusing it becomes. Your paint-over, on the other hand, is quite beautiful. Thanks for sharing

@davidrevoy I love this, David 😻

Not an artist myself, but curious about the thoughts on the "Rise of AI" in the artists' fields I did a poll not too long ago. My own thoughts were like "Gee, this is kinda worrying trend and will be sorely disruptive".

So this opposite take and embracing of what new possibilities may be in store is very refreshing to me. Thank you. You are always inspiring!

Here's the other poll with (elaborate) discussion thread:

@davidrevoy whoa! Using AI art as a helper tool is a great concept that I haven’t seen before. Nice!

@davidrevoy Beautiful paintover! :artaww: Would it have been more tedious than concepting and painting this image from scratch though?

Licensing issues aside, I am on board with using AI like any other digital tool to minimize menial work, especially when creating something like comics which tend to have a lot of repeating elements.

@davidrevoy But on the concepting side...idk. In terms of labour efficiency it might be a boon but it feels like it takes away some of the fun and can alienate you from the creative process, esp. if you obsess over it as a time-saving tool.

Art is one of those things where labour efficiency is really not the point (though capitalism loves it); taking that logic to the extreme of course you'd want to remove humans from the workflow, which imho defeats the entire meaning of art.


If we reach a point where society is consuming mostly AI art, I wouldn't even consider that a human society, and we'd probably have more things to worry about than artists going out of jobs. AI in that sense is self-defeating if it drives human artists to extinction.

On the point of using AI for concepting, do you think there is a fundamental difference between using AI to generate images vs the ability to search & draw from the unlimited amount of visual media we can find online?

@tuatarara Hey, I totally agree. I experimented a bit more, and it was fascinating at first, but it quickly felt boring. It has this 'tadaaa🎉 ' effect that is ready-made, but also never falls accurately where I wanted the art to be. Hard to really have a style or a art direction working with that unless heavily painted over. So, it is often as much trouble as starting from scratch for me right now.

@davidrevoy Yes, and it does seem that whatever style AI is able to consistently replicate the best now falls into the range of stuff that is most readily available in the data it is fed from the net. Hopefully AI art will actually compel a greater diversity of styles to flourish and be more widely recognized than the (over)abundance of sci-fi/fantasy realism we see today, as artists seek to distinguish themselves from AI!

Inscrivez-vous pour prendre part à la conversation

Le réseau social de l'avenir : pas de publicité, pas de surveillance institutionnelle, conception éthique et décentralisation ! Gardez le contrôle de vos données avec Mastodon !