|
god drat that poo poo is crazy 
|
#
?
Feb 15, 2024 21:21
|
|
|
|
| # ? May 31, 2024 20:23 |
|
|
well that got good faster than i was expecting we're not even 24 months out from Dall-E 2
|
|
#
?
Feb 15, 2024 22:22
|
|
|
Mola Yam posted:well that got good faster than i was expecting This was me a year ago, it's all going so fast that I just expect it to make huge leaps all the time now. In this instance I'm not surprised at all that text to video is so good now.
|
|
#
?
Feb 15, 2024 22:23
|
|
|
Mola Yam posted:well that got good faster than i was expecting these are only 10s or so given how other image to video (I know this is text to video) work they tend to fall apart around the 16s mark. I'd be a bit surprised if it was coherent over a minute. I'm saying this is in line with my expectations of being able to make a 15 minute silent video by end of year.
|
|
#
?
Feb 15, 2024 22:24
|
|
|
The bird video is insane, the shape and detail off the feathers don't change at all, even when it moves around. The coherence of the images is really incredible. Sam Altman is currently taking requests from Twitter users, which also look incredible https://twitter.com/sama/status/1758219575882301608
|
|
#
?
Feb 15, 2024 22:37
|
|
|
granny's got t-1000 spoon powers
|
|
#
?
Feb 15, 2024 22:39
|
|
|
e: ^^^ I swear I came up with that analogy independently This one's wild in a couple ways. I thought it looked good at first, but then the kid's hand pops up and reverse morphs itself like the T-1000 changing direction. And then I went back to rewatch the beginning - I thought the main folks were on a balcony at first, but I think the AI's getting confused and the crowd off to the side is actually just made up of tiny people 
|
|
#
?
Feb 15, 2024 22:44
|
|
|
moist banana bread posted:The first time I tried to generate something pornographic out of curiosity it output this bizarre warped mangled pile of bodies looking thing. It felt kinda like I violated the computer. It's hard not to personify this stuff and we're definitely gonna have a crowd of dudes in love with a Tamagotchi soon. I know I'm just ranting now but maybe we should be buying Hasbro stock. They just need to release an advanced enough Furby to join the bubble. I'm doing more than just running a local LLM, I'm also setting up new hardware and going over the fast.ai course so I can create my own multi-modal model to generate both text and images. https://course.fast.ai/ It's slow going but I keep making progress and I'm happy with that. I'm big on free, open source, and local run. I hope to release the model to the world when I'm done. My goal is to be able to directly ask why the image generation AI Model produced six fingers just to see what the response actually is. The answer will be a fascinating insight no matter what it says. The closest I've found so far is LLaVa which I've linked in the past. That is a text generator with machine vision. It doesn't generate images on its own but can view them and talk about them. You can try a demo yourself here https://llava.hliu.cc/. I can make a longer write-up about LLMs in general but it's honestly easier to just point you at how to run your own because it's so dead simple now. You can ask the AI itself if you want to know more. https://github.com/LostRuins/koboldcpp/releases If you're on windows there's no need for installation of anything. Download the .exe, run the program, load the model, point your browser at localhost, chat, that's it. No mess, no fuss. The new Dynamic Temperature setting is amazing too, self-adjusting models offer more than an illusion of responding to you. Especially neat in that you can partially load a larger model on both GPU and CPU at the cost of taking longer to generate. I run a 13B parameter model on a GTX 1660 Super for example, if you wanted you could run on CPU only so it's relatively accessible. Yes, LLM models can be run on some phones if you wanted to. https://huggingface.co/TheBloke is where you find LLM models that are quantized (shrunk) to fit on your hardware. GGUF files are what you want for KoboldCPP. LLM models are currently being mixed and mashed and merged together in much the same way Stable Diffusion models were early on so it's a wild wonderful chaotic mess right now. No one really knows what's best as we're all kind of throwing spaghetti at the wall to see what sticks. There are thousands of mixes for all sorts of specific fine-tunings. If you find one you like, be sure to share with the class! https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard Open LLM Leaderboard https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard Chatbot Arena Two good sources for getting a relative idea of what model to try. I can give a personal thumbs up recommendation for anything based on LLaMa 2. Mixtral 7b is also incredibly popular but I haven't tried it myself yet. Things are moving incredibly fast and anything I tell you about today could outdated by tomorrow, literally. We live in exciting times!
|
|
#
?
Feb 15, 2024 22:58
|
|
|
BoldFace posted:OpenAI announces their new text-to-video model, Sora. I need to see some otherworldly imaginative poo poo, otherwise it might as well just be a video of a loving bird. Which is simultaneously amazing and points to the fact that we've crossed a threshold, but why are these all milquetoast renders of realistic scenes when the fun happens when you bend reality and try to break the technology.  Edit: I'm scrolling what's available on the site and there are a few imaginative ones, but they're burying the lead by hiding them and showcasing the realistic poo poo imo. edit: here we loving go "A gorgeously rendered papercraft world of a coral reef, rife with colorful fish and sea creatures." https://cdn.openai.com/sora/videos/origami-undersea.mp4 XYZAB fucked around with this message at 23:18 on Feb 15, 2024 |
|
#
?
Feb 15, 2024 23:11
|
|
|
XYZAB posted:I need to see some otherworldly imaginative poo poo, otherwise it might as well just be a video of a loving bird. Which is simultaneously amazing and points to the fact that we've crossed a threshold, but why are these all milquetoast renders of realistic scenes when the fun happens when you bend reality and try to break the technology. gotta crawl before you can walk/run gotta know the rules before you can break em this post was generated by shitpostAI
|
|
#
?
Feb 15, 2024 23:18
|
|
|
Lol I'm not gonna ever find a job again
|
|
#
?
Feb 16, 2024 01:18
|
|
|
A lot more videos on their blog post. https://openai.com/research/video-generation-models-as-world-simulators
|
|
#
?
Feb 16, 2024 02:56
|
|
|
pixaal posted:Does it support Image to video so I can say seed it? When can I, a paying customer play with it? (I know this isn't public yet but I can't be the only one interested) Looks like it, follow BoldFace's link: BoldFace posted:A lot more videos on their blog post. OpenAI posted:Animating DALL·E images  https://cdn.openai.com/tmp/s/prompting_1.mp4 https://cdn.openai.com/tmp/s/prompting_3.mp4
|
|
#
?
Feb 16, 2024 03:39
|
|
|
BoldFace posted:OpenAI announces their new text-to-video model, Sora. There seems to be an impressive understanding of physics in this model. It also handles video to video and image to video. Here's an impressive image to video: https://twitter.com/angrypenguinPNG/status/1758296824484663678
|
|
#
?
Feb 16, 2024 04:13
|
|
|
Some video to video https://twitter.com/bilawalsidhu/status/1758308160488566925
|
|
#
?
Feb 16, 2024 04:16
|
|
|
those videos are nuts oh poo poo this just dropped https://www.youtube.com/watch?v=NXpdyAWLDas LifeSunDeath fucked around with this message at 04:23 on Feb 16, 2024 |
|
#
?
Feb 16, 2024 04:21
|
|
|
LifeSunDeath posted:those videos are nuts look at the legs at 2:39ish, it's amazing how it often gets walking right and then sometimes does something wacky
|
|
#
?
Feb 16, 2024 04:31
|
|
|
AARD VARKMAN posted:Lmao combining minions + sref + weird is fun as hell Solefald posted:While my probe against the horrific Minion Tits is a joke; This is a place of refined tastes and delicate sensibilities. All minions will be generated with clothing that promotes modesty from now on, so as to not invite any further unclean, impure thoughts in anyone who found themselves stirred by the forbidden allure of Minion Tits. Without further ado, here's some Burka / Hijabi Minions I made today:             Some Francisco Goya / Stephen Gammell / Junji Ito minion mash-ups         KakerMix posted:
|
|
#
?
Feb 16, 2024 05:23
|
|
|
edit: nevermind I wish human-edited videos didn't do obnoxious zoom-cuts or weird edits to cut out dead air. It's annoying, jarring, and distracting. edit 5: this guy has well-edited videos mostly. But there are times he just cuts out dead air and it fuckin sucks. Make the effort to have meaningful transition! credburn fucked around with this message at 07:36 on Feb 16, 2024 |
|
#
?
Feb 16, 2024 07:26
|
|
|
Goddamn, that's like an order of magnitude better than the ~3 months old Stable Video Diffusion. Guess those couple of billions in additional funding for OpenAI will start to show more and more
|
|
#
?
Feb 16, 2024 10:16
|
|
|
Oh man, DALL-E has no clue how to depict someone playing cat’s cradle.   It also has no idea how to depict a middle-aged transgender woman of southern Mexican ancestry in the first one. I said to do it in the style of Vermeer and it spits out a picture of girl with a pearl earring playing with what looks like a sensory toy for small children. I like the middle one though because it reminds me of trying to use a smartphone on mescaline. Especially the skin, that’s like 100% mescaline visuals, at least the way I get them. Why the gently caress do I pay monthly for this stupid thing? Oh yeah because it kills time.
|
|
#
?
Feb 16, 2024 13:42
|
|
|
some sora glitching. goddamn it mostly looks real though. ugh, this stuff was all pretty fun till these hyper realistic videos started, now I'm certain were' all gonna be enslaved. https://packaged-media.redd.it/5bus...168ae091896#t=0
|
|
#
?
Feb 16, 2024 15:30
|
|
|
Sora looks cool, but may be less useful than still images. What're the use cases for a few seconds of imperfect video, apart from meme gifs? It's still at the level where it couldn't create anything like an actual movie (no object permanence, for starters - things disappear when briefly obscured). And it's nigh impossible for a human to edit out the weirdness.
|
|
#
?
Feb 16, 2024 16:01
|
|
|
wedgie deliverer posted:Lol I'm not gonna ever find a job again I’ve never found a job in the first place
|
|
#
?
Feb 16, 2024 16:04
|
|
|
RIP Syndrome posted:Sora looks cool, but may be less useful than still images. What're the use cases for a few seconds of imperfect video, apart from meme gifs? It's still at the level where it couldn't create anything like an actual movie (no object permanence, for starters - things disappear when briefly obscured). And it's nigh impossible for a human to edit out the weirdness. This is the first step, it wasn't all that long ago that this was a good image for "Frog in a gamestop"  anyway I'm very excited to shitpost with Sora because I already shitpost with Pika and that gets me like 4 a month.
|
|
#
?
Feb 16, 2024 16:05
|
|
|
Does anyone remember the name and/or url for any of those old, primitive AIs that would generate what looks like a cluttered desk full of junk at first glance, but when you look at it closer you realize you can’t identify a single object?
|
|
#
?
Feb 16, 2024 16:19
|
|
|
RIP Syndrome posted:Sora looks cool, but may be less useful than still images. What're the use cases for a few seconds of imperfect video, apart from meme gifs? It's still at the level where it couldn't create anything like an actual movie (no object permanence, for starters - things disappear when briefly obscured). And it's nigh impossible for a human to edit out the weirdness. Obviously it's only going to get better and better. Just found this thread this morning as I never venture into GBS; those Minion titties on the other page are the most cursed things I've ever seen.
|
|
#
?
Feb 16, 2024 16:25
|
|
|
pixaal posted:This is the first step, it wasn't all that long ago that this was a good image for "Frog in a gamestop" Oh, I totally get that, and it's cool. But both the image and video generators are still in the "make image look plausible" hallucination stages. There's no executive function with any understanding of what's going on, and that seems like it'd be a more pronounced issue in moving pictures. Still image generators haven't really moved past that first step yet either. It'll happen, though. I think there've been some promising moves with long-term memory and reasoning in text generators, and some of those lessons are likely transferable.
|
|
#
?
Feb 16, 2024 16:27
|
|
|
LifeSunDeath posted:some sora glitching. goddamn it mostly looks real though. ugh, this stuff was all pretty fun till these hyper realistic videos started, now I'm certain were' all gonna be enslaved. Man, even recognizing that it's glitching out, there's a weird quality to it that's kind of convincing / not completely setting off the uncanny valley "THIS IS WRONG" effect to me when I watch it. I'd compare it to watching someone do sleight of hand magic I guess? Yes, simply pull the plastic sheet out of the hole and unfold it into a molded plastic hoverchair. This is normal. Everything is fine. Midjourney allows the word "boudoir" past the filter now, and so I did the natural thing and made a bunch of boudoir photos of vegetables. Prompts are variations on the below: brick walls, a(n) [vegetable], paint me like one of your french girls, in the style of oil paint on canvas, boudoir photography, grungy, the [vegetable] is lightly misted with water, the [vegetable] is tangled in bedsheets --ar 3:2 --c 50 --w 1000            A couple oddities that came out of the set:    vegetables that came out slightly different when I tried similar prompts in Bing / DALL-E   
|
|
#
?
Feb 16, 2024 16:35
|
|
|
Nigmaetcetera posted:Does anyone remember the name and/or url for any of those old, primitive AIs that would generate what looks like a cluttered desk full of junk at first glance, but when you look at it closer you realize you can’t identify a single object? Art breeder still has rhe older ones last i saw
|
|
#
?
Feb 16, 2024 16:37
|
|
|
RIP Syndrome posted:Sora looks cool, but may be less useful than still images. What're the use cases for a few seconds of imperfect video, apart from meme gifs? It's still at the level where it couldn't create anything like an actual movie (no object permanence, for starters - things disappear when briefly obscured). And it's nigh impossible for a human to edit out the weirdness. On demand "stock" gifs for web content? Improve your chatGPT generated article on how to eat a pineapple with funny gifs of pineapples! There are rumors going around in CEO circles that those drat genZ kids of today don't like to read pictures anymore and only want to see moving thing. Also, as someone else already said: that's only like 1-2 years of progress from Lovecraftian nightmare visions of Will-Smith-Thing eating pizza. And they got infinite funding now, which they hadn't before.
|
|
#
?
Feb 16, 2024 16:41
|
|
|
Pre-DALL-E image gens were hot poo poo. Made these in 2019 or 2020, don't remember how: "trump fursona"  "trump fursona drinking honey polaroid"
|
|
#
?
Feb 16, 2024 17:02
|
|
|
I was using those sorts of images for my CYOA back in the day, I found they stimulated a lot of imagination. You don't need to do nearly the mental filling in of blanks with modern AI images, which is fun to look at but it doesn't inspire me to write anything weird like it used to.
|
|
#
?
Feb 16, 2024 17:09
|
|
|
Sab669 posted:
If you play with the fire you'll see far more cursed. I have attempted to make Danny DeVito kicking a football as a test image with controlnet poser one of my first. I'm very bad at human anatomy. SD decided he looked more like he was taking a poo poo so it made a wrinkly rear end birthing a football. So cursed.
|
|
#
?
Feb 16, 2024 17:17
|
|
|
            
|
|
#
?
Feb 16, 2024 17:20
|
|
|
Soulhunter posted:
hell yeah put them titties on the glass, carrot.
|
|
#
?
Feb 16, 2024 17:41
|
|
|
Finally, I feel seen. https://www.youtube.com/watch?v=mcYl70vq_Ns
|
|
#
?
Feb 16, 2024 17:55
|
|
|
LifeSunDeath posted:those videos are nuts He points out that in the Tokyo video the reflections move at a different framerate than the rest of the scene, and I guess this might mean that they self-produce some of the training data by just recording random fly-throughs of scenes in Unreal Engine or Unity. I gotta revise my prediction for the future a little bit - I previously have said that the future would be billions of insect sized camera drones flying around the world, because the all-knowing AI is drowning in so much generated poo poo that it has to use drones it controls to know what the real world looks like. Now my prediction is the AI will use the billions of drones to generate training data for OpenAI.
|
|
#
?
Feb 16, 2024 18:04
|
|
|
ymgve posted:He points out that in the Tokyo video the reflections move at a different framerate than the rest of the scene, and I guess this might mean that they self-produce some of the training data by just recording random fly-throughs of scenes in Unreal Engine or Unity. https://twitter.com/ralphbrooks/status/1758230974700130445
|
|
#
?
Feb 16, 2024 19:25
|
|
|
|
| # ? May 31, 2024 20:23 |
|
|
pixaal posted:This is the first step, it wasn't all that long ago that this was a good image for "Frog in a gamestop" thats still a pretty good image of a frog in a gamestop imo 
|
|
#
?
Feb 16, 2024 19:49
|
|
