I am holding the phone in one hand, and holding my papers in the other hand!

These vids make my day and break the sadness and depression! 🥳

and it's actually two minutes long for once

The program is ‘text-to-image’ and yet it still sucks at texts! Why trust OpenAI at all when we create our own hundreds of folds better than what we are seeing on paper?

Some folks prefer the unit PPS (puss per sec)

Need another factor of 1000 so you can peta-cat!

The imperial unit is furballs per fortnite

I, for one, embrace the catularity

as a cat i a prove of this

As an AI developed by Google, I strongly encourage you to only post inclusive comments. There is nothing unusual about years. All years should be considered equal. Please only post responsible comments. Thank you for being on the team.

The roller coaster has taken a sudden turn 😂

@@itskittyme What.

XDDDDD​@@itskittyme

Probably every single year from now on will be unusual

a word paints 1000 pictures

❌FPS 😸CPS

Leapfrog for sure. As the checkpoint creators aren't at gunpoint obligated to feed only "legally approved" images for training data and has far more controls for the user.

Exactly! I love SD for this reason.

Yeah, but he gets a lot more engagement from up-to-the-minute material. It's hard to position yourself as bite-size tech news if you aren't going to give people the bites that are actually making public waves, even if they're weak stories.

@@michaelleue7594 yeah, I'm sure. Now that he's gone full time, he can't afford to just cater to the nerds

It’s difficult when so much news comes out so quickly

Well I think the biggest problem is that this _is_ the most important tech news, but Dr. Zsonai-Feher is not an expert on this subject and the mechanics of the subject are far beyond even summary, when information on the tech is available at all.

@@SeventhSolar he used to do a decent job of summarising the innovations and architecture of new models though. It was clear he had read the papers.

Strongly agree

a hero walking on artist corpses

With pineapple

yes

As an salami crab, I have no say that I want to be free.

AI taking off is bad by the way

I'm kinda gonna miss the spelling mistakes and six fingers, ngl

but we can put the bottle inside the genie

@@strangelaw6384Thanks, Valentino.

why would you put it back, its an amazing genie

As I see it the biggest problem with SD isn't image quality but understanding the prompt. The next big leap forward would thus come not from improvements in the diffusion model itself but from upgrading or replacing the CLIP, perhaps using some kind of intermediate scene description language to gain a better control of how different objects relate to one another.

@@clovernacknime6984 That's the nice thing about having LoRA, ControlNets and masks, you have more precise control. Will be very nice to have the prompt simply understood better of course, but having that extra control adds layers of possibility.

Modeling the human mind without considering the suffering that could be simulated unintentionally. Interesting thought experiment.

Our only salvation... Depp Learning.

Things like that already exist for SD1.5/SDXL (IP Adapter, Instant-ID, etc.); they each use a ControlNet architecture. I'm not sure if ControlNet can be applied to the new architecture, but I'm sure if it gets popular, researchers and devs will find a way to make it work 😁

they were 2 minutes long videos too

This channel still does a solid deal of good discussing highly underrated papers every month or so

​@@zrakonthekrakon494, true

Humans are actually ai

there's also duplicate fingers. people use it to know if they're dreaming

@@CharldeonWait, duplicate fingers on other people or yourself? When I’ve had lucid dreams my hands are perfectly represented down to the smallest pore and wrinkle.

@@leejerrett8268 it really varies. i had 20 lucid dreams and there was like 2 times where it was too realistic my hands didn't look malformed

@@Charldeon That’s interesting, I’ll have to keep an eye out for that. I once had a lucid dream where the dream characters I was with tried to convince me I wasn’t dreaming by telling me ‘you can’t be dreaming; look at how detailed your hands are’ almost as if they knew hands weren’t something they usually bother to render fully so they put in a bit more work filling in the visual and tactile detail on my hands just to try to pull me back into the illusion.

of course, you can create a LoRA, search it up

Consistent character views is possible but still kinda messy to do on diffusion models, you can search stable diffusion character sheet or turnaround and see what you can find, also there's a embedding called CharTurnerV2 that seems to help with it.

@@chiguirolover Great, thank you so much. 👍

@@sa8212 np, good luck

Not as far as I'm aware; I think that's just AnimateDiff applied to the results.

iirc automatic1111 has a plugin for generating a turn around sheet

stabillety has a model caled stable video. it can make a image into a video

Part of the charm!

now squeeze those papers!

It's been announced, but not released yet. There will be an open beta of sorts accessible by a waitlist you can get to through Stability's website, but it will be under noncommercial license terms and will likely be remote rather than running on your computer (they've mentioned getting access to a Discord server, so it will probably be like the testing periods for SD1 and SDXL). I don't expect the "open beta" to start for a couple more weeks, but that's a guess.

@@Adreitz7 how is this considered "open source" then? A complete misuse of the term. It's utter bullcrap

@@arisoda Open source doesn't mean "give me all your things right now". Just like SD1 and SDXL, there is a development phase, a (NC, remote) testing phase, and then a following open release of both the code and the models. This is how StabilityAI operates. It's been done before, so there is no reason to freak out.

@@Adreitz7 oh

VRAM is not more expensive than normal ram. It's just that Nvidia are greedy, greedy pigs who are intentionally nerfing the VRAM amounts on their Geforce cards so that companies won't get any silly ideas like paying them less money by just buying a bunch of GTX4090 cards instead of the 20x more expensive datacenter cards.

Well, remember, our visual cortex does a lot more than just connect concepts with images. It also does object recognition, visual memory, motion estimation, spatial location estimation, control of the eyes and arms during reaching or focusing, among many other things. So it's not a fair comparison to begin with 😁

@@IceMetalPunk I know it's quite different and the sophistication of our visual cortex is vastly greater than this but it wouldn't be able to faithfully produce images of cats if it didn't have equivalent visual understanding of a cat, indeed generation is much tougher than perception. Compare your reading speed on a given topic to your writing speed of equivalent quality material. Indeed our brains can't do what these image generators do. We cannot produce mental images of this quality level. Our dreams are more high level abstractions. Sort of semantic level image content rather than bothering with the actual pixels and light distribution or fluid dynamics. And to be fair the brain is in continuous adaptive operation. All ML so far is static systems. Both static topology and weights. And yes, turning eyes is not something they handle yet or keeping track of coherent surroundings under body movement. But I still think it says that a quite capable visual cortex can be quite small. And that's also born out in biology as little hamsters can perform quite well with a much smaller brain. Birds too. Not to mention classical algo visual cortex statically implemented in a custom chip. That might be crazy lean and fast. Vision doesn't need to change really.

Stable diffusion is open source and can be used for any purpose and ofcourse there is no watermark so even if they restrict you can use it, they will never gonna you generated it from their platform

No he just talks like this.

no more evil youtube brainrot

Dibs on the idea of custom porn based on the user with custom dream girl :D