I just found this:
http://blog.rhino3d.com/2018/06/new-machine-learning-examples-with.html
https://blog.bricsys.com/the-potential-of-ai-for-bricscad/
... this is a hidden competition than everyone is missing?
... FreeCAD or Blender have any advances in this field?
... should we worry?
@jtm2020hyo said:
... should we worry?
No more than the rest of the people in this world...
I played with the grasshopper tools a bit back then when they came out, also with Crow and Galapagos. The main issue with all these cool sounding technologies is not integrating them, but finding something useful for them to do in our field.
Machine learning is actually pretty simple. The problem is, you need a massive amounts of data on a particular problem, then you can teach the "ai" and let it run on new cases and benefit from the speed increase. This makes uses like autocorrect, translation and image recognition or underwear removal quite obvious. But unless we really want to let the machine spit out individual houses autonomously, as I suggested in another thread, its usefulness in the design process seems to be very limited. (I've seen some studies on urban planning and very specific areas like traffic infrastructure)
Of course if you have any ideas and care to share them, it is very welcome.
I was checking some GPT-3 videos, is awesome, and scare me...
...as you mentioned , looks like AI building design is still green...
...I just could suggest create a workflow taking an (hypothetical) upgrade version of to FreeCAD + KiCAD, I mean (CSV+TXT)+XML+IFC , or (Worksheet Tables+Definitions)+Diagrams+3D modeling. Until now this workflow is the most powerful than I know... And is not just for circuits or MEP, Architecture and Structural design can be benefited, some examples:
Blender and spatial diagrams:
AutoCAD Electrical diagrams and Inventor 3D assembly
KiCAD digrams and assembly and FreeCAD 3D modeling
... I mean give human intelligence and automated processes to our projects.
You should obtain your key from OpenAI for GPT-3 :)
By the way I assume that you are all aware of the security issues with this, but just as a reminder, a little rant.
There is a reason why all the major companies are pushing ai technologies other than marketing. To make the most of the technology, the data has to be processed online. You don't have to care that Huawei scans and stores all your photos and everything you do with your phone (I do), but if Autodesk releases ai on the 90% US offices using Revit, they could have a lot of fun. I'm not saying they are willing to, for example, fiddle with international competitions, but there are other, even bigger powers in this world.
I see a potential in evolutionary algorithms for now though, if anyone is also interested in this, sverchok should have a working GA node:
Clearly buildings can be designed with an evolutionary approach, though my suspicion is that it only really works with building forms that are 'adaptive' in the first place. i.e evolving traditional architecture is possible, but modern architecture I'm not so sure, how would the computer know if the result was any good?
I don't see AI and machine learning (as distinct from genetic algorithms) designing buildings any time soon.
[Edit] I think AI will be very good at writing marketing texts for architects.
@JanF looks interesting
Also, it seems that Sverchok has a great UI improvement, and I hope its performance improves too, it uses a lot of CPU in my laptop
These days I see more and more that people in other industries say: "We want solutions that don't touch our data"
@brunopostle NVidia and Autodesk work on this area and (except a small mistake they have) they're so close to that point to design buildings and its interior automatically
They use raster plans to generate (bubble and rectangular) graphs and fields (quantities) as input for machine learning
Then use that data/information/knowledge as input for generative processes
Also, use some correction and optimization methods like Bayesian methods
@ReD_CoDE said:
@brunopostle NVidia and Autodesk work on this area and (except a small mistake they have) they're so close to that point to design buildings and its interior automatically
But are these good buildings?
This is their old version:
https://developer.nvidia.com/blog/archigan-generative-stack-apartment-building-design/
It depends, I like the approach they follow
But I have another workflow in mind, which is different and will take time to complete it
@ReD_CoDE said:
This is their old version:
https://developer.nvidia.com/blog/archigan-generative-stack-apartment-building-design/
These are all plausible looking plans, in the same way that gpt-3 generates plausible sounding text without any understanding of what it is doing. I don't see how you get from here to something anyone would consider building.
It is very cool that a computer can do these things, it's like a dog that can walk on its hind legs.
Imagine soon (not soon, until 2050) AI can be a mathematician that solves questions that are impossible human mathematicians solve
And even this happened decades before! For instance, computers found that we can explain all color ranges (2D) with only 4 hues. Before that 5 color-map was popular
So, when computers can solve impossible questions in mathematics and physics, do you think they can't learn how to design?
All of them are mathematics, even architecture
@ReD_CoDE said:
So, when computers can solve impossible questions in mathematics and physics, do you think they can't learn how to design?
All of them are mathematics, even architecture
I know computers can design buildings, I just think that machine learning and AI are unpromising approaches. I agree that maybe in 2050 things will look very different.
Sverchok has genetic Algorythm node if something
ArchiCAD from v22 has user behaviour prediction ML algo, but it trying to guess and it takes time and if it guess wrong, than you waiting for his extra-job
that nvidia link look awesome, I wish play with something similar in FreeCAD or Blender.
... in the next group video call it would be very interesting for them to talk about this topic and their progress, and if possible share some samples.
Bumping this thread on the strength of what I've been seeing happening with AI text to image solutions like Disco Diffusion, Midjourney and DALL.E in recent weeks. Like all things new fangled, there are issues, but generally speaking, the output is mind blowing, communities are building around these tools and 'AI whispering' is becoming a thing, a skill set even, for both professional concept artists and digital art hobbyists. Some even find it therapeutic. Some are getting bored, others are just getting started, there are philosophical debates going on, millions of surreal images flooding the internet, graphic novels being made on the fly, people losing sleep, text prompts being shared and remixed ad nauseam. A lot of stuff can improve, and some are already improving, like new improved trained models being released at greater speed (I think for example that a UI system needs to be created for dialling weights and testing variations without losing previous iterations, Artbreeder does this to a reasonable extent), but really, the horses have bolted out of the stable and are off at a fast trot.
And then, there is architecture.
I've seen a few architects have a go at things, and what works for concept art works fairly well for architecture as well, in terms of iterating through visual concepts quite quickly with the help of AI and narrowing down / optimising for a preferred look and feel which can be futher developed with other tools. The output in not half bad to be honest.
Now, an open source alternative to DD, MJ and Dall.E, called Stable Diffusion just gave out beta access (now closed), so there is a FOSS option. I shudder to think of what sorcery could come out of jamming Stable Diffusion into a cauldron already filled with great smelling stuff like Topologic, Homemaker, BlenderBIM, Archipack, et al and see what cooks from the mix. I see potential for these trained AI models being optimized for architecture and being able to take spatial blobs from Topologic for example and automate what Homemaker does in skinning it with a stylized envelope of choice, with dials for adjusting weights in one direction or the other. However, what happens when AI gets it's dirty fingers on the accompanying BIM metadata? Would I for example be able to peg my project budget within a certain threshold and have the AI run cost optimization on the fly and then spit out variations that fit the current cost band, while still respecting other weights I may have set?
It is my utmost wish that no one suffers sleepless nights on account of all this, but if they do, I'm all here to see what they come up with, and learn what prompts / hacks they deployed, ditto all words of caution and hesitation that are bound to come! ;0)
@DADA_universe said:
Bumping this thread on the strength of what I've been seeing happening with AI text to image solutions like Disco Diffusion, Midjourney and DALL.E in recent weeks. Like all things new fangled, there are issues, but generally speaking, the output is mind blowing, communities are building around these tools and 'AI whispering' is becoming a thing, a skill set even, for both professional concept artists and digital art hobbyists. Some even find it therapeutic. Some are getting bored, others are just getting started, there are philosophical debates going on, millions of surreal images flooding the internet, graphic novels being made on the fly, people losing sleep, text prompts being shared and remixed ad nauseam. A lot of stuff can improve, and some are already improving, like new improved trained models being released at greater speed (I think for example that a UI system needs to be created for dialling weights and testing variations without losing previous iterations, Artbreeder does this to a reasonable extent), but really, the horses have bolted out of the stable and are off at a fast trot.
And then, there is architecture.
I've seen a few architects have a go at things, and what works for concept art works fairly well for architecture as well, in terms of iterating through visual concepts quite quickly with the help of AI and narrowing down / optimising for a preferred look and feel which can be futher developed with other tools. The output in not half bad to be honest.
Now, an open source alternative to DD, MJ and Dall.E, called Stable Diffusion just gave out beta access (now closed), so there is a FOSS option. I shudder to think of what sorcery could come out of jamming Stable Diffusion into a cauldron already filled with great smelling stuff like Topologic, Homemaker, BlenderBIM, Archipack, et al and see what cooks from the mix. I see potential for these trained AI models being optimized for architecture and being able to take spatial blobs from Topologic for example and automate what Homemaker does in skinning it with a stylized envelope of choice, with dials for adjusting weights in one direction or the other. However, what happens when AI gets it's dirty fingers on the accompanying BIM metadata? Would I for example be able to peg my project budget within a certain threshold and have the AI run cost optimization on the fly and then spit out variations that fit the current cost band, while still respecting other weights I may have set?
It is my utmost wish that no one suffers sleepless nights on account of all this, but if they do, I'm all here to see what they come up with, and learn what prompts / hacks they deployed, ditto all words of caution and hesitation that are bound to come! ;0)
Just found someone already trying things out:
Application for Stable Diffusion Beta has been opened again (gets swamped pretty quickly). Heads up to anyone who might be looking to get in: https://stability.ai/beta-signup-form
@DADA_universe said:
I shudder to think of what sorcery could come out of jamming Stable Diffusion into a cauldron already filled with great smelling stuff like Topologic, Homemaker, BlenderBIM, Archipack, et al and see what cooks from the mix.
It's definitely an interesting field! but I don't think a text-to-image Diffusion Model could be directly useful for the AEC industry. In fact, I don't believe that on the short term, it's likely that we see good results on any Supervised Learning approach within our field (it's difficult to have large amounts of standardized data). On the other hand, some Reinforcement Learning sorcery applied on a very robust environment could be the way to go. But of course, that's hard, and requires a lot of experimentation with the exploration logic.
I have been deep diving in MidJourney for the past 2 weeks or so, if I may bring a bit of insight concerning architecture. This generation of AI models is not capable of creation. It can spit out very believable re-interpretation of known sources, and it can even spit out astounding crossovers that would even make the executives of Marvel salivate. It has been taught flawlessly all the concepts of design, color, composition, golden ratios, but it is not capable of creation. You can make the case that all art currents and all artists have been directly or indirectly inspired by what came before them, but I do not see this generation of AIs replace the "creative" part of the job of architects, especially because it is so intricately tied to the actual technical needs of every project. It can spit out amazing things, but in order to do that, users have to learn and master a new language, very similarly to what coders here are doing with open source projects.
Here's a few renders of what would have been a collaboration between Zaha Hadid and Le Corbusier :
!
!
!
!
!
!
Also, there is the ever-growing subject of training data biases. Will it be useful to people in South Asia if the training set only contains apartment plans of Northern American buildings ? How will the trainer be able to gather enough data to train it AI efficiently ? And for what purpose ?
It will IMO 100% be used as a tool to make the productivity of the users skyrocket, same as the drawing board in its time, same as the typewriter in its time, same as the personal computer in its time, same as the access to internet in its time. It will 100% be integrated into most softwares that we will be using in a couple of years (look at AI filters in photoshop !), either buried deep in the code so the end-user doesn't even know it's there or out there in the open to help with publicity stunts.
I'm eagerly waiting for the next generation of "AIs", as I am waiting for the next generation of laser scanning devices, and the next generation of photogrammetry software. And drawing tablets. And Quantum Computers. It will change how we work, and we need to try to predict how if we want to take advantage of it, or at least not let our jobs become obsolete by then. Who draws their technical plans with a pen on paper today ?
Maybe we should start tot think how all this could affect open source software?
This could very well spell the end for OSS in my opinion...
I thought this was the idea of Autodesk's 'One' program. It's in the cloud. Your designs aren't shared with others (unless you want to share them) but it 'sees' everybody's CAD work and learns from them. The AI will have a massive source from which to learn. The AI then gives the user suggestions about how to improve what they are working on.
Login or Register to reply.
