Glad someone seems to have made a more presentable app for this. I'm barely capable when it comes to downloading Git files and running them (not a programmer).
The FAQ is pretty scarce, though, and I wonder if it will be able to be updated as the developers change the model.
There are a dozen of providers for SD and other models to run for fractions of a penny with decent ui/ux needing 0 programmer knowledge, but prompt knowledge is growing.
HuggingFace
PromptHero
come to mind, but many can also just be opened in a colab notebook
I guess you have to download it to see how much "better" the UI is; I couldn't find any evidence of that on the App Store or the developer's site.
And by "non-native" I guess they mean that DiffusionBee is wrapping non-Swift code, or what? It certainly feels like a (pretty simple) native app, and for what it does I find the UI plenty good enough.
I was going to install this but it requires macOS 13.1, and I try to stay one version behind Apple's "improvements" so I'll have to wait.
nobody has any info but given its from Sindre i’d put a small amount of money on this using the Swift native reimplementation from Apple rather than the original python one ported to Apple Silicon with an unknown bag of tricks.
but its just a guess, i have no proof; entirety of my reasoning is this wouldnt be sufficiently different from diffusionbee to work on if this were not the case
I can't use the App in the OP since it's for M1 only and I have an Intel Mac. But I appreciate the pointer to DiffusionBee! Also, this is the first time I've seen the author's website - and he has some nice apps available! Definitely going to play with some of these.
This author has very high opinions on what he thinks a better UI is. He has some other app, where he sells his application better because it has better UI than mine.
If you want to contribute to something like this, but that’s open-source, I started working on a similar app - SwiftUI + Apple’s CoreML Stable Diffusion here: https://github.com/justjake/Gauss
I’m currently working on downloadable models before I publish to the App Store & notarized builds to GitHub. A download manager adds complexity but should cut the initial app install size to ~50mb, and then distribute models via GitHub Releases.
any idea if CoreML Stable Diffusion is appreciably faster or slower than regular-SD-on-Python?
( sorry if this is somehow well known, ive been trying to pay attention but havent seen anyone comment on what effect the implementation difference has had on the system performance)
Hey, thanks for doing this. I am looking around your git repo.
I wrote a Swift book with a couple of CoreML examples. Apple did a fine job supporting deep learning models incorporated in macOS, iOS, and iPadOS apps.
Generally I use a libre license like GPLv3 for end-user apps to prevent others from quickly flipping them for profit, and open source license like MIT for libraries.
If you need to release a closed-source fork I’m happy to sell you a custom license.
I also made an app that allows to generate images using CoreML (with help of the apple library) on iOS and macOS (using iOS) https://loshadki.app/imagegeneratorml/ very basic and easy app
Doesn't seem to work very well. I just tried to generate an image with default settings. A progress bar starts to show some work. Once it gets past around 50%, an image starts to form and gets a bit better as it goes on. Before it reaches 100%, the image is gone and the whole process starts over.
I left it for about 15 min this way. Then, started DiffusionBee while this app was still trying to render my image. I copy-pasted the image prompt into DB, set it to the same 30 steps and it rendered an image that resembles my prompt a lot better than the partial one from Amazing AI in less than a minute.
Amazing AI is still on the same loop. I'm killing it now.
36 comments
[ 3.4 ms ] story [ 78.8 ms ] threadThe FAQ is pretty scarce, though, and I wonder if it will be able to be updated as the developers change the model.
HuggingFace
PromptHero
come to mind, but many can also just be opened in a colab notebook
> I don't plan to support this. Check out DiffusionBee if you need more functionality (if you can live with worse UI and a non-native app).
According to the link itself, it has fewer features but better UI than DiffusionBee.
And by "non-native" I guess they mean that DiffusionBee is wrapping non-Swift code, or what? It certainly feels like a (pretty simple) native app, and for what it does I find the UI plenty good enough.
I was going to install this but it requires macOS 13.1, and I try to stay one version behind Apple's "improvements" so I'll have to wait.
It’s a nice app though!
but its just a guess, i have no proof; entirety of my reasoning is this wouldnt be sufficiently different from diffusionbee to work on if this were not the case
Are there customization to be done to obtain good results ?
I’m currently working on downloadable models before I publish to the App Store & notarized builds to GitHub. A download manager adds complexity but should cut the initial app install size to ~50mb, and then distribute models via GitHub Releases.
( sorry if this is somehow well known, ive been trying to pay attention but havent seen anyone comment on what effect the implementation difference has had on the system performance)
https://arstechnica.com/information-technology/2022/12/apple...
Sure, an M1 Ultra computer costs 3.6K€ higher than an RTX 3060 computer nowadays, but still, it gets close to the best performance on the market.
I wrote a Swift book with a couple of CoreML examples. Apple did a fine job supporting deep learning models incorporated in macOS, iOS, and iPadOS apps.
Would you consider changing the license from GPLv3 to the ISC license?
https://choosealicense.com/licenses/isc/
If you need to release a closed-source fork I’m happy to sell you a custom license.
https://machinelearning.apple.com/research/stable-diffusion-...
https://github.com/AUTOMATIC1111/stable-diffusion-webui is also popular but requires a lot more config to get going.
I left it for about 15 min this way. Then, started DiffusionBee while this app was still trying to render my image. I copy-pasted the image prompt into DB, set it to the same 30 steps and it rendered an image that resembles my prompt a lot better than the partial one from Amazing AI in less than a minute.
Amazing AI is still on the same loop. I'm killing it now.