This group is dedicated to the experimental AI Visualizer for Archicad 27, which will reach End of Service on December 31, 2024.
The new AI Visualizer in Archicad 28 is now fully cloud-based, with advanced controls like creativity and upscaling. Click here for further details.
- last edited on by
Laszlo Nagy
I have downloaded the visualizer as per the instruction on my pc, however the server does not generate the image. Could you help with this.
Adam
Similar problem, though no issue connecting to AI server in my case.
However I suspect the problem in my case may be related to slow response time of server
All 4 checks above passed.
Looking at the terminal window it seems like a process is still ongoing, with no errors reported there, even as ArchiCAD reports a non specific error, advising the user to look at the terminal window for more information.
Given the extreme slowness evident in the progress bar indicators in the terminal window (on even default AI image settings in the ArchiCAD AI dialog) I am guessing this is some sort of timeout issue?
If this speed is representative (around 20 minutes just to get to 25% of first iteration, before ArchiCAD gives up), even if an image were to eventually result (i.e. if ArchiCAD relaxed its timeout limit), I can't see this being a particularly useful tool.
I had already seen that ArchiCAD reporting an error does not halt the generation process as reported in the terminal window, but interestingly even hitting the 'Restart AI Engine' button in the ArchiCAD AI dialog does nothing at all either - the process reported in terminal window continues uninterrupted.
Any attempt to generate another image while the the current terminal window is still active (even after the reset button is clicked) triggers another non specific error message, with same referral to the terminal window, which continues to work away on the first image and which still shows no error.
Almost as if the ArchiCAD is just not communicating at all with the processes it initiates
Only way to proceed is to manually close the terminal window and click 'Start AI engine' to open a new one.
Of course this also eventually fails/times out in the same way.
Hi GAG, my RTX3080 has 10GB (full spec below) - so is well above minimum requirements, if not at the full 16GB recommended.
Noting also that I have tried running via the low VRAM version batchfile, with no change in performance/outcome
*** Edit *** if I drop the number of iterations below the default 80 to the minimum 20, I do get an outcome before ArchiCAD times out, though it takes several minutes rather than seconds. The outcome not especially useful, although that is a separate issue & not the fault of the tool itself.
Given the speed, I would probably do better to upload a basic screenshot from ArchiCAD standard 3D window directly to one of the online AI generators to work with rather than than trying to process it locally.
Obviously with the current state of the art, I would get tend to get results that don't withstand much scrutiny architecturally either way, except under very constrained circumstances or after very many trial and error iterations with various prompts.
Certainly thus far am not seeing anything about this tool that would justify the staggering cost of a GPU upgrade to a next gen card (ironically it seems to be all the AI farms out there that are gobbling up every available current gen GPU and driving up prices for everyone else).
@Paul King wrote:
Hi GAG, my RTX3080 has 10GB (full spec below) - so is well above minimum requirements, if not at the full 16GB recommended.
…
Certainly thus far am not seeing anything about this tool that would justify the staggering cost of a GPU upgrade to a next gen card
Hi, Paul.
Thank you for posting the full GPU-Z report as a screenshot, not just video card name, this helped me to find a possible workaround and a way to keep your money (at least for a while).
Before we move on to a possible solution, I must warn you that, despite my thorough research into your video card's features and compatibility, you will need to double-check all my conclusions!
So, the answer is: Resizable BAR.
I quote NVIDIA's news about Resizable BAR support in the GeForce RTX 30 Series:
Resizable BAR utilizes an advanced feature of PCI Express to increase performance in certain games. As of March 30th, 2021, Resizable BAR is supported for GeForce RTX 30 Series graphics cards.
For desktops to take advantage of Resizable BAR, users need a GeForce RTX 30 Series graphics card with a supported VBIOS, a compatible CPU, compatible motherboard, motherboard SBIOS update, and our newest GeForce Game Ready driver.
Resizable BAR is an optional PCI Express interface technology. As you move through a world in a game, GPU memory (VRAM) constantly transfers textures, shaders and geometry via many small CPU to GPU transfers.
With the ever-growing size of modern game assets, this results in a lot of transfers. Using Resizable BAR, assets can instead be requested as-needed and sent in full, so the CPU can efficiently access the entire frame buffer. And if multiple requests are made, transfers can occur concurrently, rather than queuing.
End of quote.
When you read about game's textures, VRAM and GPU transfers, you could easily extrapolate this to machine learning and Stable Diffusion.
What we can notice on your screen?
Resizable BAR: disabled.
So, the solution is to perform your video card software upgrade to get a performance boost.
Thanks GS - that is a very interesting suggestion and very helpful research!
After a bit of a struggle was able to update the GPU firmware (noting that neither the Gigabyte Control Centre firmware update function method nor the manual bios update executable method actually work by default with Gigabyte RTX GPU firmware updates - a long standing issue apparently. A workaround for anyone interested is posted here)
A couple of motherboard BIOS settings needed to change as well, but end result is Resizable Bar is now active.
While still slow, this time same settings did not time out completely on lowest resolution image in High Vram mode, but they still do time out on highest resolution in high Vram mode - even using minimum iterations settings (20)
In low VRAM mode, while still slow (still maybe a minute or two), it is noticeably faster with 20 iterations, on highest listed resolution (1216x832).
In low VRAM mode, with highest resolution, and highest available iterations (50), image generation does complete, and takes around 3.5 minutes
That said, while I can see the AI is riffing with the scanned shapes from ArchiCAD and applying its large catalogue of statistical patterns related to images of buildings, the lack of any innate understanding of how a building works is pretty evident.
This is still far from being useful as a design tool, and you would never want to show the results to a client, however nicely rendered - the basic and obvious mistakes might well get them questioning your fees.
Essentially this is generating a stylistic composition that has a superficial resemblance to a building without really reflecting any spatial organizational substance or functional purpose - the exact opposite of the actual architectural design process we are paid to undertake.
We are not short of much more coherent glossy magazine images already if we are just looking for inspiration
Lower prompt strength seems to generate slightly more plausible results (assuming you have enough of the critical elements resolved and modelled yourself).
But prompt strength set to the minimum 25%, gives results that are still far from usable
And of course having the critical elements modelled already then brings in to question the need for AI at all at that point, except to the extent I guess that you could create the impression of having done more modelling entourage work than you actually have.
But the minute the client asks you to extend a bedroom while leaving everything else as is, or to show you the floor plan corresponding to the image, the whole thing is exposed as fake, since without a resolved underlying model with 1:1 relationship to the image, these things can't (yet) be done, as far as I can tell.
I suspect the real value of AI as a design tool would kick in when AI can generate 3D models reflecting some insight as to the way buildings actually work - though hopefully (from a gainful employment point of view) that moment will not arrive too soon!
