🧵 3D DALLE called Point-E
Anonymous at Tue, 20 Dec 2022 17:35:33 UTC No. 930374
It isn’t as good as Google’s DreamFusion, but blazing fast! Like ~600x faster to generate . 2D DALLE has already turned the creative world upside down. How will 3D DALLE disrupt games, VR, metaverse, …?
Bonus: it’s fully open-source! Both the code and all model weights (for 40M, 300M, 1B parameters) are available at https://github.com/openai/point-e
Arxiv paper: https://arxiv.org/abs/2212.08751
2023 will be a year of powerful 3D generative models, and 2D will simply be their planar projection.
It's truly over
Anonymous at Tue, 20 Dec 2022 19:31:48 UTC No. 930386
Well I guess it could kill some applications of 3d in the future, say, where the end result is a render like archviz or product visualization, assuming you can crank the output resolution high enough to not see individual points, but I can't see it "killing" 3d graphics in general. It's limited to points? Points/voxels obviously suck because they're slow and difficult to work with. We all want polygons, don't we? I suppose you could retopo volume to mesh but having sculpt-level resolution for something like a coffee cup? Unless there's an AI that can fully automatically retopo stuff sensibly I don't see it working at all.
I wonder how they trained the models though. Photogrammetry? Sounds like it'd take a shitload of photos to train even the most basic stuff.
Also is a retopo/uv AI feasible at all? I don't really understand how these "AI" work. I get the impression that it's not as simple as the oversimplification currently thrown about by buttblasted artists that it just collages parts of similar pictures. It's just too good to be just that. But I can't see how a pure retopo/uv AI can be trained at all. Maybe you could get a ton of humans to somehow "tag" most common patterns and evaluate their quality, and it could somehow extrapolate from that, but it's still fuzzy because artists make topology for good normal smoothing, good vertex data interpolation such as uvs (I suppose that includes normals as well) and maybe good deformation, except cutting corners for speed, convenience or efficiency, and I can't see an AI understanding all those factors without being an actual "AGI".
Anonymous at Sun, 1 Jan 2023 16:44:41 UTC No. 931750
>>930374
WARNING Blog Post INBOUND !
3D AI art.
There are multiple aspects who often exclude or conflict with one another.
>ASPECT 3D VS 2D.
Lets be real 2D is fake it fakes perspective of the 3D world.
In principle giving 3D models to an AI and telling it that this is a spoon or a screw driver makes the AI understand the actual 3D geometry of the object.Instead of trying to make the dumb thing eat 100s of pictures in hopes of it figuring out what a screw driver looks like.
>simple 3D and AI art
Another related subject is using simple generic 3D rigs and posing them in 3D space and telling the AI
>This pose rig is superman
This transmits incredible amounts of information to the AI unlike a stupid text box. The AI now knows the rotation and position of humanoid limbs in relation to one another.
It placing something that looks like a IRL superman over this rig is a great trick.
However
>AI is theft
AI is theft 2D AI only exists because it has stolen 100s of 100s of pictures artists made, human artists.
The joke is that 3D models are rare and not so posted on the internet this means that there are not a lot of 3D models the AI can steal to understand 3D objects.
>AI is a scam
Understanding the previous point AI is a scam. the AI can literally never exit the pictures it was trained on. It is literally not better then a search engine who gives you only the pictures it was given to its database.
Not sure what is so great about getting a 3D model some other human made.
And knowing this the 3D AI will be unable to steal enough 3D models.
What OP posted looks like some toddler made these 3D models.
Anonymous at Sun, 1 Jan 2023 16:51:43 UTC No. 931751
>>930386
>get the impression that it's not as simple as the oversimplification currently thrown about by buttblasted artists
It literally is only this.
It literally only recalls what it was given.
See picture it literally forces everything into this anime style with the same neo-Victorian architecture because it literally can not understand more. You can look into this, people tried to break this chink AI by submitting very un sexy teen faces.
Yet the AI can only do these 3 faces and will force everything into it. Obviously it was given enough human art to save to its database or they did not have enough to steal from human artists.
Play around with this AI and see for yourself.
h5.tu.qq.com
Anonymous at Sun, 1 Jan 2023 20:56:19 UTC No. 931791
>>930374
https://www.youtube.com/watch?v=uNW
Anonymous at Mon, 2 Jan 2023 16:51:23 UTC No. 931885
>>931791
technology is so cool
I really hope some anti-tech movement doesn't start and stop our progress toward the future.
Anonymous at Mon, 2 Jan 2023 19:53:31 UTC No. 931904
>>931884
kek
Anonymous at Mon, 2 Jan 2023 22:38:46 UTC No. 931918
>>930374
why do these computer dorks so desperately want to automate human expression? do they lack ambition that badly?
Anonymous at Tue, 3 Jan 2023 01:57:37 UTC No. 931930
>>931751
Thats just the chink AI that probably uses less then 500 dataset images. It is better to compare it with something like Midjurney or Stable Diffusion, that do give you basically everything. The AI can connect the dots between things it learns. It can take images of men with glasses, women without glasses, men without glasses, and then from this information give you women with glasses, or at least this experiment was performed 5 or so years ago with this tech.
>>931918
They are trying to automate literally everything. For them it is the question of not if we want to automate it, but that we must be the first ones to automate it because some other people will do it instead if we wont do it now. Humans are slaves to progress, creative jobs are being automated first by AI because they have large margin of error, and because there is lot of data for it to learn. Lawyers and doctors could theoretically be automated right now, but because the machine does mistakes 10% of time, it is unsuitable for said job, even if the human counterpart does worse mistakes 20% of times. It is not them trying to attack creative jobs, it is them attacking all jobs at once and seeing which ones get fucked over. Lot of the AI researchers say they would never want to have children because they fear the future this much. Humanity will march on progress no matter who is doing it, so being against AI to them seems like futile endeavor, even if all AI researchers agreed AI is bad idea and will only cause damage, they will still continue researching it.
Anonymous at Tue, 3 Jan 2023 02:01:08 UTC No. 931931
>>931791
Looks like less control shittier version of what Blender bone extrusion would do, at most saving 20% of the time. You still need to put the bones on the joints, you still need to find the hands and fingerbones, and so on. It just looks shit.
Anonymous at Tue, 3 Jan 2023 02:09:14 UTC No. 931932
>>931930
>For them it is the question of not if we want to automate it, but that we must be the first ones to automate it because some other people will do it instead if we wont do it now
"they" cannot even remove bagging and cashiering jobs at the grocery store
Anonymous at Tue, 3 Jan 2023 03:51:47 UTC No. 931938
>>931932
Because those jobs do not require purely cognitive skills. AI is AI, artificial intelligence. It does the thinking. Computer that does the thinking will not pack your fucking groceries. For that you need robotics, which are far behind AI research, since unlike AI robotics actually require smart people doing all the researching and engineering. With AI you only need to make the pipeline, which is most of the time pretty simple, and then just feed it lots of data. AI research proved itself that just taking what is essentially just random neuron spaghetti and throwing any sort of info from the internet at it makes it smarter, just throwing as much shit from the internet at it makes it create itself. This is why AI research is going so fast right now, because it is super simple to do, and the only limiting factors are data and computing power. Only fraction of the internet has been used for machine learning, and there are special chips meant for running AIs being developed, chips that are supposedly able to make AT training 100 times cheaper. 2022 was just the beginning of machine learning. And yes, it will not replace wedgie jobs, since wedgie jobs require robotics to advance faster.
Anonymous at Tue, 3 Jan 2023 05:25:53 UTC No. 931941
>>931918
most people are genuinely incapable of differentiating between "soulless product" and art. proof: virtually all music, video games, movies and art of this millennium.
i mean pic related is considered/known as one of the "greatest" current "artists". i recently paid 39.99 for a 4-pack of his shitty-fake-art-garbage socks so I can symbolically wear it on my smelly manfeet (just like egyptian pharaohs used to walk on pictures of niggers that were painted on their soles).
Anonymous at Tue, 3 Jan 2023 14:05:35 UTC No. 931964
>>930374
2009 Roblox meshes be like
Anonymous at Tue, 3 Jan 2023 14:53:20 UTC No. 931972
>>931938
>And yes, it will not replace wedgie jobs, since wedgie jobs require robotics to advance faster.
anon, they've been developing robotics for god knows how long. Boston Dynamics has been a thing for god knows how long. They aren't replacing shit and things like SD cant be used for commercial purposes
Anonymous at Tue, 3 Jan 2023 16:28:32 UTC No. 931986
>>931964
50,000,000 daily active users be like