(A word of warning to anyone thinking they might take five minutes to install this: be prepared for it to take a long time and use a lot of bandwidth. It’s not just that the whole thing is tens of GB big, it’s that it requires dozens of Python modules and their dependencies, and it deliberately disables caching and ignores already-installed packages. So it downloads, for example, torch (226MB) dozens of times, once for every package that depends on it. I shudder to think why it does that, or what the total download size is, but it’s been saturating my 20Mb/s link for about an hour and a half, and it’s nowhere near halfway through the list.)
I realise that doing this yourself maybe be out of reach for you, either hardware wise, or technically. Which is why I made this offer, no payment required; I just love doing this so much, and I’d love to see what gets created from other peoples ideas.
As a way of expressing the magic that I feel from this, check this out:
You may know of the artist Michael Whelan, he’s done lots of album covers
I created this VCV logo with the prompt An arcane machine with lights and wires by Michael Whelan
Edit: don’t be fooled by the speed of the video, I have an RTX 2080 Ti, it gets very hot during the process, which for this image took 7m 55s. This image is 750 iterations at a resolution of 512, then for the still that was scaled up x2 using ruDALL-E Real-ESRGAN, the video was super slow-mo’ed by a factor of 20
Another cool technique I’ve discovered is using a feedback loop, where you generate an image, then use that image as the seed image for the same prompt, and continue that iteration to see where you end up. Depending on the “likeness to seed” settings, this can go off in really interesting directions.
I used the technique on this set of abstract Spiderman images:
Installation went pretty straightforward, I had two errors in the pip installer scripts, but that seems to be just uninstallers not finding the right version to uninstall, so I ignored that. The “Machine Learning Test and Activation” inside Visions of Chaos was positive. Now Downloading the ML-Scripts, but that seems to be super slow (around 1MB/s), so I think I will leave it running overnight.
The Fractal Flames of the non-ML-modes are my favorite so far. I already have a good amount of new desktop-wallpapers
Will try movie-creation next… what an awesome piece of software, thanks again for sharing.
Come across another app on instagram, I’ve not used this one, but might be an option for some, I think it allows some free usage, not sure if there is a catch or not
Well you get 5 credits for free, but a lot of the options need extra credits, such as increasing the iterations or resolution
It does seem fairly fast and has most of the options of the Visions of Chaos app and the CoLab notebooks, but clearly the credit system does not lend itself well to the trial and error approach, I can see this cost adding up quickly
On the other hand it is just a website, so is a very low barrier to getting a quick hands on test of the text-to-image concept
A text prompt of A metallic topopolis in a rainbow nebular with a super nova sun gave me this output:
With just a little bit of trial and error in Visions of Chaos, I evolved the above prompt to produce this (in my opinion, more aesthetically pleasing) output:
Thanks Dan for bringing Visions of Chaos to our attention! I find it way easier to use and a lot less flaky than the collab notebooks. Everything is in one place, including detailed instructions for getting it running. And the software is deeply packed with a billion other juicy options for generating cool imagery.
I used the first screen grab as the seed image and played around with various text prompts that all included some form of “eurorack” & “electroluminescent wire”. I had to negatively weight “wood:-10” to tone down the wood panels that kept showing up. Haha. Following the seed image are some of the better results.
For me, Text-to-Image always produces either an all-black image or an error message along the lines of this one:
RuntimeError: CUDA out of memory. Tried to allocate 36.00 MiB (GPU 0; 6.00 GiB total capacity; 3.82 GiB already allocated; 0 bytes free; 3.96 GiB reserved in total by PyTorch)
Is my installation broken, or is my 6GB graphics card really not enough to run this? Before I run it, Windows tell me that there’s only a few hundred MB of GPU memory in use.
From what I have read: this is the most common error and basically means your GPU is not good enough for the settings you have selected.
Not sure, but I would guess that CUDA allocated memory is only part of the process, or the currently executing step, not representative of the memory required for the whole operation.
First thing to try is making your output resolution smaller.
I normally render at either 256 then super res x4 or at 512 and then super res x2
I’ve also read that Cutouts and Cutout batches have impact on the memory use, but I havent played with them much.
You could also try unchecking one of the models ViT-B/16 for example
Sorry, no its not super enough, specifically its not an RTX, therefore has no CUDA cores.
The machine learning stuff needs an RTX 20x or 30x Nvidia card
Great idea, I wanted to see what sort of material was in the models for eurorack so I whipped up a
batch of quick colour perlin noise init renders for the prompt eurorack module photographs to see some ideas:
)!
