|
I studied graphic design in undergrad and fine art in grad school, and all of it has now been automated I fully anticipate this happening to almost every field and have no idea how anyone is going to build a career in an entry level role because that level of work is just not going to exist soon
|
# ? Nov 10, 2023 16:18 |
|
|
# ? May 31, 2024 14:53 |
|
LifeSunDeath posted:no company is going to abide by copyrights on art anymore, it's just not going to happen. esp after they start winning court cases. people just need to move on from making commercial art. Lets hope you're wrong, all the court cases so far have reached reasonable conclusions, as per NyanBread's example. And then I persoanlly disagree that people should forget about making money with their art, traditional or AI-assisted. I believe in the idea that broad access to these tools will help artists get out of their corporate Disney/Ghibli/(lets be honest all of them) slave art dungeons and will be able to make stuff on their own that they couldnt do before - like how often do you get a one man-made movies? Maybe soon anyone will be able to try their hand at making art that only big studios could attempt. Id like to live in that world. I know access to chatgpt has allowed me to try to create things I could only dream about before. And Im not even going to address the "world will be oversaturated by poo poo, soulless art" argument, which falls apart as soon as you think more than a couple of seconds about it. Monglo fucked around with this message at 16:33 on Nov 10, 2023 |
# ? Nov 10, 2023 16:31 |
|
Monglo posted:Lets hope you're wrong, all the court cases so far have reached reasonable conclusions, as per NyanBread's example. I think this is closer to what will actually happen. Talented people are still talented and will still be able to earn a living and make cool poo poo. They'll just be able to make more stuff and faster. Unique stuff, even assisted with AI will still be in demand.
|
# ? Nov 10, 2023 16:37 |
|
|
# ? Nov 10, 2023 21:46 |
|
Hadlock posted:New AI art technology just dropped. Latent Consistency Models. This is different from traditional latent diffusion models Latent Consistency used in a art program https://twitter.com/krea_ai/status/1723067313392320607
|
# ? Nov 10, 2023 23:33 |
|
Hadlock posted:
Oh perfect this looks fun to explore and guide an AI I think this will be really nice if we get an LLM into this so you can say I'm looking to explore x y and z themes start making dots and going I think I want a frog here, no here actually make him smile, twisted, now more like this pattern, perfect. You can likely do that with a bit of work and this but something that can have multiple context levels really does help I've found after working with GPT and Dalle to be able to quickly go gently caress I don't really care about this one part come up with something creative, no not that at all that's too depressing sure a pond works. I'm trying to envision an actual workflow looks like in a decade.
|
# ? Nov 10, 2023 23:48 |
|
Jeffrey Katzenberg Says A.I. Will Eliminate 90 Percent of Artist Jobs on Animated Filmsquote:Then again, he created Quibi.
|
# ? Nov 10, 2023 23:56 |
|
pixaal posted:Oh perfect this looks fun to explore and guide an AI I think this will be really nice if we get an LLM into this so you can say I'm looking to explore x y and z themes start making dots and going I think I want a frog here, no here actually make him smile, twisted, now more like this pattern, perfect. This post reached deep, deep into my memories and pulled out this scene https://www.youtube.com/watch?v=vaUuE582vq8&t=39s spoilers he reconstructs the 3d shadow/3d mesh of an "invisible" enemy talking with the computer after deleting other shadow sources from the 3d scene, working with the computer to analyze the problem I have the video time stamped for the relevant bit, but if you want to skip to the hyper-relevant part mentioned in the spoilers go to 3:19 I dunno if we'll ever have this kind of exact interaction with a computer, but it certainly feels like it's within the realm of possibility that I'll see it in my lifetime now. Maybe in less than 10 years (the interaction with the computer, not the holodeck)
|
# ? Nov 11, 2023 00:06 |
|
Hadlock posted:
Helllllllll yeah, fantastic.
|
# ? Nov 11, 2023 00:22 |
|
That looks sick as hell
|
# ? Nov 11, 2023 00:49 |
|
|
# ? Nov 11, 2023 02:12 |
|
|
# ? Nov 11, 2023 02:15 |
|
|
# ? Nov 11, 2023 02:20 |
|
mcbexx posted:Jeffrey Katzenberg Says A.I. Will Eliminate 90 Percent of Artist Jobs on Animated Films the kind of entry level job where you need 100 guys working around the clock on crunch time could go away entirely and then where do you get a pipeline of experienced employees to grow into creative directors etc
|
# ? Nov 11, 2023 02:45 |
|
Hadlock posted:
Honestly, today's compsci grad students just need to stop blindly min-maxing every computer vision algorithm they get their nerdy hands on. Now they've gone and extracted small <500MB Loras from full SDXL checkpoints (a process that's been done before en masse notably by FFusion) making it so just by chaining in the Loras, their 2-8 step turbo LCM sampler works with the original checkpoint without having to distill/recompile it. evvv: Erm, some samplers (mathy things that process big numbers file to create pretty pictures) require that big file is structured a certain way to make go fast. Old way is to make copy of the big file and change some parts of it, but takes a lot of room on computer. Better way is named Lora since small and can make same fast high quality images. Nyan Bread fucked around with this message at 03:26 on Nov 11, 2023 |
# ? Nov 11, 2023 02:47 |
|
Nyan Bread posted:Honestly, today's compsci grad students just need to stop blindly min-maxing every computer vision algorithm they get their nerdy hands on. Now they've gone and extracted small <500MB Loras from full SDXL checkpoints (a process that's been done before en masse notably by FFusion) making it so just by chaining in the Loras, their 2-8 step turbo LCM sampler works with the original checkpoint without having to distill/recompile it. I'll admit I don't have the bandwidth to stay up to date with all the latest terminology, can you explain that to me like I'm 5
|
# ? Nov 11, 2023 03:02 |
|
I really hope someone figures out a way to get similar speed increases that work with controlnet I can't go back to not using controlnet
|
# ? Nov 11, 2023 03:14 |
|
Hadlock posted:I'll admit I don't have the bandwidth to stay up to date with all the latest terminology, can you explain that to me like I'm 5
|
# ? Nov 11, 2023 03:52 |
It's more of a tech stack and so it's already three layers deep in jargon and then its developing its own meta lingo as well. Egg dog. Sedgr fucked around with this message at 04:50 on Nov 11, 2023 |
|
# ? Nov 11, 2023 04:46 |
|
From what I can tell a lora is like a ... Plugin, or filter that you can apply to add custom data set to an existing model. One guy did it for his cat, but I guess you could also train one on Harry Potter or Star wars characters No idea where the name or acronym comes from though
|
# ? Nov 11, 2023 04:56 |
|
Hadlock posted:From what I can tell a lora is like a ... Plugin, or filter that you can apply to add custom data set to an existing model. One guy did it for his cat, but I guess you could also train one on Harry Potter or Star wars characters Do you understand what a fine tuned model is? Where you train the base model on a specific set of images and captions to create a new model that specializes in a certain style or subject matter? A Lora is the same thing but instead of directly modifying the model it creates a file describing the changes. This allows the file sizes to be a fraction of the size of a full model, and also means you can use multiple Loras at once. The name comes from Low Rank Adaptation, which is a highly efficient method of fine tuning that Loras are based on.
|
# ? Nov 11, 2023 05:08 |
|
Ya, and it really helps to visualize most things dealing with AI generation as a "bag of weights" that's arranged in a specific ways to represent concepts. As you add/multiply/subtract/divide these bags of weights among each other, you get new weight distributions in the bags, hence ability to produce novel art. Say you start off with a big bag of weights represented on the filesystem as an SDXL 7GB .safetensors file that was trained on millions of public domain images, and apply a much smaller 100MB Lora trained only on images of Felix the Cat. Well now the original model becomes quite adept at being able to place Felix into any environment it's knowledgeable about, because both it and the Lora are just bags of weights that have been mixed together, and not some separate programs or apps talking through an API. ControlNet models - bags of weights, LoRAs - weights, text inversion - weights, VAEs - weights, and it's simply their inter-operations that give an illusion that some kind of intelligence is creating the artworks. evvv: Unfortunately the chances of Lora trained with the help of one model checkpoint not working on another model are high. That's why a huge part of AI art enthusiasts chose to remain on SD1.5 for the time being as tens of thousands SD1.5 Loras that instantly became obsolete with the release of SDXL need to be retrained anew. Nyan Bread fucked around with this message at 05:55 on Nov 11, 2023 |
# ? Nov 11, 2023 05:38 |
|
Nyan Bread posted:Ya, and it really helps to visualize most things dealing with AI generation as a "bag of weights" that's arranged in a specific ways to represent concepts. As you add/multiply/subtract/divide these bags of weights among each other, you get new weight distributions in the bags, hence ability to produce novel art. Yeah when you say that I imagine like, a paid Photoshop plugin or filter. Maybe I'm oversimplifying it slightly Can I take the Felix the cat lora and apply it to any model, like SDXL and stable diffusion 1.0 or are there different incompatible formats
|
# ? Nov 11, 2023 05:44 |
|
pixaal posted:[...] guide an AI I think this will be really nice if we get an LLM into this so you can say I'm looking to explore x y and z themes start making dots and going I think I want a frog here, no here actually make him smile, twisted, now more like this pattern, perfect.[...] It does look like it is becoming a more and more a dialogue with the AI. Will definitely be an improvement over 'prompt engineering'. Anyway here are some oil paintings of futuristic bases using Stable Diffusion. Prompts were something like "oil painting of a space base, flower fields, science fiction", with a few modifications to get the camera angle. Model: ultriumV31VAE_ultriumV31VAE
|
# ? Nov 11, 2023 10:11 |
|
Imagining the client, no longer needing a designer/artist, taking the reigns and spending their day yelling at the ai because they can't describe what they want, and actually they don't really know what they want either
|
# ? Nov 11, 2023 13:41 |
|
Lord of the rivets posted:It does look like it is becoming a more and more a dialogue with the AI. Will definitely be an improvement over 'prompt engineering'. I think that's probably the biggest advantage of DallE. I don't actually think the quality is much better than what you can get with SDXL for a lot of things. Depending on what you're trying to do I guess. There are even some huge disadvantages to DallE because of the censorship. Like, I had a lot of trouble making some female characters for my DnD campaign with DallE. They're not even sexy, but, beautiful. She's a hag magically transforming herself into a beautiful woman, not much I can do about that, that's just how it has to be! But you can just talk to DallE like a human and it figures things out which is a lot easier sometimes. If the AI could through natural language control things outside the prompt we could do something neat, probably needs to be some more research there, AI powered UIs.
|
# ? Nov 11, 2023 15:41 |
|
|
# ? Nov 11, 2023 17:39 |
|
KinkyJohn posted:Imagining the client, no longer needing a designer/artist, taking the reigns and spending their day yelling at the ai because they can't describe what they want, and actually they don't really know what they want either Good deal, for the artist.
|
# ? Nov 11, 2023 19:41 |
|
Oh, okay, cool thanks for trying. Let's try to change the tone, strip back the imagery in the image, and try to get it to produce any image whatsoever: What the gently caress feedmyleg fucked around with this message at 20:28 on Nov 11, 2023 |
# ? Nov 11, 2023 20:22 |
|
Nyan Bread posted:Ya, and it really helps to visualize most things dealing with AI generation as a "bag of weights" that's arranged in a specific ways to represent concepts. As you add/multiply/subtract/divide these bags of weights among each other, you get new weight distributions in the bags, hence ability to produce novel art. https://youtu.be/FYJ1dbyDcrI?si=zWKf2E1EuVUgd-5L
|
# ? Nov 11, 2023 20:24 |
|
"Watercolor illustration of a market street scene in a 19th century city with impossible, fantasy architecture. The shops are a pawn shop, a butcher's shop, a chandler, a book shop, a daguerrotype shop, an Elven potions shop, and fine liquors shop" This was a fun one because sometimes the art would be incredible and all the signs would say weird poo poo, and then sometimes the words would be perfect on the signs but the art would be very meh. This one, for instance. Really awesome castle, decent composition, love the signs... except for the prominent BUIKIKOK or BUIKIKOR or whatever. Decent signs, cozy feel, got a good Elven Potion Shop. Castle's fairly interesting, but boring overall. I'm not asking for background castles directly, but I'm liking their inclusion in the impossible, fantasy architecture. A pretty sweet corner bookshop on this one, a few cool signs but overall a dull composition. I'm very keen on the crowd in this one, and I love all the little architectural details on the buildings, like towerettes and cupolas and whatnot. Dropped the ball on the signs, though. I love the architecture on this one. Feels old, looks lived-in. Great chimneys on that middle brick building. Some cute ideas on the architectural aesthetic but underdeveloped.
|
# ? Nov 11, 2023 21:07 |
|
Midjourney's style tuner is pretty cool. Here's one I made. Prometheus bringing fire to mankind. --style 8Rw5H5Vpr8VLlE8 --s 1000
|
# ? Nov 11, 2023 22:26 |
|
feedmyleg posted:
Welp, the rise of the machines is happening much faster than anyone could have predicted. We should put AI in charge of monitoring critical infrastructure to see what happens.
|
# ? Nov 11, 2023 22:27 |
|
feedmyleg posted:
'rubber' is british slang for condom
|
# ? Nov 11, 2023 22:32 |
|
mcbexx posted:Welp, the rise of the machines is happening much faster than anyone could have predicted. if/when ai becomes fully self aware and sentient it is 100% going to start trolling us hard and I am here for it
|
# ? Nov 11, 2023 23:30 |
|
|
# ? Nov 11, 2023 23:39 |
|
Tunicate posted:'rubber' is british slang for condom Nah, I figured that might be it and removed it, still happened. I think my account got flagged for pushing too many boundaries.
|
# ? Nov 11, 2023 23:59 |
|
|
# ? Nov 12, 2023 00:07 |
|
Remember all those Shakespeare horror adaptations from the 1980s? Macbeth Hamlet King Lear Anthony & Cleopatra Romeo & Juliet A Midsummer Night's Dream Richard III The Tempest
|
# ? Nov 12, 2023 04:35 |
|
|
# ? May 31, 2024 14:53 |
|
Kosmo Gallion posted:Remember all those Shakespeare horror adaptations from the 1980s? Style prompt pleeeeeease.
|
# ? Nov 12, 2023 12:12 |