#dall-e 2

[sequelanet]

Dall-E 2: incrível inteligência artificial que transforma textos em imagens - Blog Sequelanet

O DALL-E 2 (openai.com/dall-e-2/) é um programa de inteligência artificial que cria imagens a partir de descrições textuais.

[valentinofmannara]

Gioielleria e anelli da donna creati con intelligenza artificiale - Parte 1

View On WordPress

[ericfiori]

Visions & Memories, by Blake Wood

[pavel-nosok]

OpenAI Announces Call for Experts to Join its Red Teaming Network

OpenAI is launching its Red Teaming Network, seeking domain experts to improve AI safety measures. The network aims to collaborate with professionals from diverse fields to evaluate and “red team” AI systems. The network will provide insights into risk assessment and mitigation, with members contributing varying amounts of time. OpenAI values diverse expertise and encourages applications from…

View On WordPress

[noeyecontactpleasse]

I'm so glad that generative AI is dying, but please please please understand that chatbot AIs and art generation AIs and voice replicating AIs are not the ones I'm scared of, they're just the most gimmicky, and the easiest to fight in a court of law.

You should be concerned with facial recognition software being used in policing. Many of these AIs are trained on datasets of majority white males, meaning they become less accurate and less effective when used on someone outside of those demographics, meaning more innocent people of marginalized communities being accused and harassed by police.

You should be concerned with resume-reading AI. Amazon created an AI model that was trained on previous hiring data and used to filter out applications. The software recognized a pattern of applicants with non-white names being rejected and thus filtered out every applicant with any non-white name. This software was deemed a failure by Amazon and was scrapped, but highlights a big issue with AI in general. If there's a shred of bias in a testing dataset, that bias will be amplified by the software.

You should be concerned with autonomous weaponry used in warfare. See, when I say "I want fewer people to die in war", it's because I want less war. What I don't want is for large wealthy imperialist countries to be putting lethal weapons with little to no human oversight into warzones where they can 'identify enemy combatants' and 'eliminate targets' at the discretion of a computer program. You know it's bad when even Elon Musk has publicly supported a ban on autonomous weaponry.

You should be concerned with AIs being trained to diagnose medical problems. These AIs are being trained on diagnosis data of previous medical cases. Meaning using patient data for a testing dataset (HIPAA/PIPEDA/GDPR etc. violations??). Anyone who's been tossed around the medical system could tell you that removing the humanity from healthcare will only lead to more suffering.

These are only a few examples I know off the top of my head as someone studying computer science in undergrad. This isn't to say that all AI is bad. AI is such a huge category of software that lumping it all together is irresponsible. AI is used in maps and navigation, spellcheck, translation software, speech-to-text, text-to-speech, online banking, malware detection, hate speech filtering, and so many assistive technologies for disabled people.

I'm glad that chatGPT and AI art generators are dying, but they're barely even on my list of concerns when it comes to AI technology.

[davidcreator]

DALL-E 2: Criação de Imagens

Você está em: Início > Artigos > Inteligência Artificial > DALL-E 2: Criação de Imagens Olá! Caro leitor, este artigo é para quem esta procurando tecnologias de inteligência artificial para ajudar nos trabalhos do dia a dia A criatividade é um elemento fundamental da expressão humana, e a evolução da inteligência artificial (IA) tem ampliado os horizontes da criação visual. O DALL-E 2 é um…

View On WordPress

[leam1983]

On Generative AI as a Process

ChatGPT: You're basically slotting coins into one of those old Mechanical Seers you'd sometimes see at county fairs or on boardwalks, back in the fifties. The difference is the Seer claims to know exactly what you're statistically determined to want to hear, but is otherwise a beautiful fount of ignorance hidden behind self-effacing verbiage. The worst part is all of this somehow was enough to blow the Turing test out of the water, even if you're made very much aware that there's nothing going on in there if no prompt is being processed.

DALL-E 2: There's rumors of a DALL-E 3 incoming at OpenAI and even more niche whispering of an insider version of the project that is wholly and completely uncensored - and it's likely to put Midjourney to task, apparently. As of now, however, using DALL-E 2 feels like its name is an indication of where its strengths lie. Are you looking for a Cubist rendering of Michael Fassbender riding a motorcycle out of the Milky Way? It's got you covered. If, on the other hand, you're trying to push for photorealism or for the stylistic trappings typical of Optane renders, you're likely to be a little disappointed by the results. It's a painter, no ifs, ands or buts - and its understanding of other visual art forms still leaves much to be desired.

Midjourney: you're handing off commissions to a genius painter and illustrator that can emulate almost any style, but that still has a relatably frustrating time comprehending the physics of the human body. Seeing it try and fail to produce believable human hands is so true-to-life you're starting to wonder why nobody's tried to come up with a picture-based alternate of the Turing test. Unlike a painter and more like an idiot savant, however, it conflates concepts together that shouldn't be agglomerated and makes weird assumptions based on your prompt.

It almost feels Taoist, in a sense. Is there a difference between someone yawning openly and someone screaming, honestly? If not, then what's the real frontier between belting out a yawn and thrusting out your chest to push a sustained note as an opera singer? A human knows there's degrees to it all - but a machine?

Stable Diffusion: unlike the other two, SD feels like a community shop where you're free to pick up tools to do whatever you want, whenever you want. The model itself exists in demo-grade and constricted forms online, and it's also the basis for all the Anime face generators that now pollute app stores - but its real power shows when you take it home and spend at least one full afternoon wrestling with Git, Python and Hugging Face credentials to turn your humble gaming rig of choice into your own personal digital Pygmalion.

There's a catch, though: earlier Stable Diffusion models used to solicit your GPU to the maximum of its abilities, which I've seen toast graphics cards within weeks. The latest revision takes a wiser approach, but it dramatically changes the way you can approach generative art.

As of SDXL 1.0, slotting in a prompt feels like setting a cake to bake. It used to be you'd be guaranteed a ho-hum result within five minutes, after hearing your GPU's fans scream for their life. Now, if you do what I did for my first run and give it the max amount of passes at 1028x1028 pixels, you're looking at a nice and cozy experience that actually lets you use your PC for other things while it's working - at the expense of time.

I asked for four simple images at the best generation parameters possible, and it took my home-based Stable Diffusion webUI eight hours to render anything - with xformers appended, too!

As of now, then, using Stable Diffusion is like baking a long recipie. It stands neck-and-neck with Midjourney, is fully uncensored and entirely under my control - but getting anywhere costs an entire evening. If you're a fan of pushing for variation after variation on Midjourney until you get something perfect, you'll be here for a long fucking while.

All the other ones: BlueWillow, Pepper AI, all the modded Stable Diffusion WebUI installs with a bespoke URL, an account and credit system, all those vaunted open platforms for NSFW generative content - they're all a waste of time, honestly. Text-based systems don't factually need to be able to crank out verbal smut, not considering the type of use they're getting. Image-based ones are running off of deprecated Stable Diffusion imageweights and typically heavily bias towards an expected use. If your need falls within their purview, then you're in luck. If all you want is big-breasted Anime waifus, you don't even need to bother with running Stable Diffusion natively. If, like me, you're using generative art to mock up WebUI elements or generate partial elements for promotional campaigns, running natively is invaluable.

On the whole, both Midjourney and Stable Diffusion have been invaluable to me for pursuits both personal and professional, seeing as both of them clearly state their product cannot be covered under copyright, which leaves their intended use as purely illustrative. Coming up with complete mockups within minutes was instrumental for our Web design team, seeing as I could quickly communicate the sort of design presence the company was shooting for and just as quickly receive revisions or refinements to said mockups until we'd be in agreement. At that point, generative art left the premises entirely and the Web Design team got to work from scratch.

Privately, running stat sheets for NPCs for a D&D campaign and being able to quickly gin up a portrait made it easier to reach my desired tone, in the understanding that if I ever want to give my characters or concepts their just desserts, I can turn these images' basic prompts into someting verbally elegant, and just send that off to my favourite commissions-friendly artist to get that still-elusive human touch. Similarly, getting unique desktop wallpapers out of what's essentially Art Class shitposting is always fun, but I can always commission them for something I could take to a scanner and use as a desktop wallpaper.

Closing Thoughts

Knowing that all that I've had these systems output can't be used to make a single buck, legally, is a sobering and comforting thought - and it makes the now-storied excesses of the early days of visual generative AI all the more disturbing. Nobody should win art gallery prizes with a piece designed by a GAN, and artists shouldn't have to fear that their likenesses and voice are at risk of being exploited. I shouldn't have to fear using both of these for exploratory projects - but I also have the responsibility to not identify as an artist and to not claim ownership of the resulting text blocks or images. Considering, it makes the hucksters on Amazon selling booklets on prompt-work for ChatGPT to shit out entire novels in a few weeks feel a little tactless - as if their entire industry wasn't at risk of collapsing the exact minute a few savvy lawmakers get their act together.

Plus, if you metaphorically consider commission work to be a sort of human-to-human prompt engineering, eight hours for SDXL to produce four images doesn't seem like such a bad deal - at the expense of the process. I haven't asked for commissions often in the past, but there's something fun to knowing that you're at the top of an artist's pile for the next several days, and could even receive emails containing said artist's interrogations, their need for added refinement, their suggestions - all in service of something that'll a) put some food on their table and b) that matters to you.

Considering, I'm not entirely afraid for the wider aspect of human-led creative purposes.

[xavierlewis]

Do you know what is DALL-E 2 and how to use this for generating images? DALL-E 2 is a powerful tool for generating images from text descriptions. It's easy to use and can provide valuable creative inspiration and visual content for a wide range of applications, from marketing and advertising to design and art.

Let's learn how to use dall e 2 in detail!