InvokeAI (an alternative to webui)

[illtellyoulater]

Looking at the youtube video below InvokeAI seems to feature a neat UI and a really swift workflow for various scenarios (inpainting, outpainting, infinite canvas, etc.)

Has anyone already tried it? How does it compare to webui ?

https://www.youtube.com/watch?v=hIYBfDtKaus
https://github.com/invoke-ai/InvokeAI

[bugmelone]

I like automatic1111 webui, it's more about experimenting and researching, finding the best way to generate something (promt, formula, recipe & etc). It so many features in one place, and sometimes it feels like i'm a kid in sandbox, i have so many ways for inspiration!
But for more artistic workflow i really like InvokeAI too. It easy for quick understanding, easy to concentrate. I'm still using both, and switch them in some cases.

After upgrading to 2.2, now i love invoke even more because today I tried its new infinity canvas feature on my graphical tablet. And it was really cool. Looking forward to implementing the node system for more customization options.

[Kilvoctu]

I gave invokeAI a try, and I wish automatic1111 webui's interface was as slick and neat as that one.
That said, my main impression is that invokeAI is missing a lot of features and/or their documentation is not great.
For example, there seems to be an API, but the docs on it is three sentences, one of which is telling me to dig through a 1k+ line python file for more information 😕
Don't want to make a list, but invokeAI is missing enough stuff to where I personally am not interested in using it yet.

Sidenote, invokeAI also is ~5GB larger than a1111 webui. Not counting any models and such.
Close to 20GB total, and it doesn't have SD2.0 stuff yet. Why is it so massive 👀

[illtellyoulater]

For what I could see so far from the videos the most important features are there, only they have been streamlined in a way that makes them less intrusive and more natural to use.

For example when you are in the Unified Canvas view, depending on where you place the bounding box, the system will decide automatically if it's going to do inpainting, outpainting, generate another image, or a mix of those (if the bounding box overlaps one or more images!).

Now that they also released a one click installer I'll definitely give it a try it asap, but if it works as expected then I think it's going to be a game changer considering also the UI.

[Kilvoctu]

The most important features to me personally is an API or otherwise some way to use it from a mobile device (similar to --listen or --share in a1111 webui). I also like using a lot of the random scripts/extensions that the community here has developed.

I did watch the video and installed and used invokeAI before making my post. As mentioned, I do think invokeAI's UI is very slick and neat but will wait until they add the features I want.

[illtellyoulater]

@Kilvoctu not sure what --listen or --share specifically do, but both WebUI and InvokeAI are implemented in a client-server architecture so exposing InvokeAI to other hosts in your local network or even the internet should be easily done and not much different from WebUI... have you checked the command line options at https://invoke-ai.github.io/InvokeAI/features/CLI/ ?

[Kilvoctu]

Those two argument are documented here

I did look at https://invoke-ai.github.io/InvokeAI/features/CLI/ and somehow missed the "--web--" argument. I guess at the time I was more focused on the API section, hehe. Thanks, I'll try it again later. 👍

[ClaraSanders]

The moment Invoke adds model merging I will probably switch to them especially now that I know that the merged checkpoints that Automatic1111 produces are faulty, ask the guys at InvokeAI. They had to fix a bug that makes Automatic1111 merged models unloadable in other any stable diffusion repo, great right?

[EricBintner]

I love InvokeAI. It's what I run at home. It's missing things like Deforum and Dreambooth, but I don't use those in Automatic1111 because they're either broken or poorly designed and impossible to use (dreambooth in the most recent Auto1111 version is useless) so I'm sticking with Google Colab for those anyway. Plus Invoke runs beautifully on my Mac Studio so no contest.

The only reason I don't use InvokeAI at work is because it's got a React UI and is impossible to use in a shared environment.

[illtellyoulater]

I agree... in the current state of A1111 there are too many things that are either undocumented or just don't work as expected.
Personally I'm getting tired of this approach of introducing as many features as possible just for the sake of that and then having to spend hours reading discussion threads on here, on reddit, discord, 4chan, and what not just to make sense of how those things work or to come to the conclusion that they are not actually working...

[Inosss]

I tried invoke ai. It looks really cool with the idea of drawing on a new canvas. Brushstrokes, brushes, variations layered on the image with the option to choose from batch images, selection of the inpainting area. It's really cool. But then everything is crooked somehow. I would like to see features for image editing in img2img from invoke ai at A1111

[gwbyrd]

The videos I saw of InvokeAi look very slick, but it seems to be missing all of the extensions and scripts in A1111.... I feel like I'd love to use the InvokeAI UI, but I rely heavily on the scripts, particularly the XY Plotting, Randomize, and Magic Prompts.

[illtellyoulater]

I finally had the time to try InvokeAI and I must say the Unified Canvas approach is really something... Of course extensions are missing but it already comes by default with many useful features (see https://invoke-ai.github.io/InvokeAI/#invokeai-features) and according to the devs many more are going to be added soon. I'll keep an eye on it... in any case competition is always good!

[Rogal80]

So, I did a comparison of SD and IA, same prompt, seed, high-res fix
Maybe I am doing something wrong?
this is AUTO1111

Invoke AI

Any clue about that?

[Inosss]

why it shoud be same?

[Rogal80]

Also same model and still AUTO is matching closer to prompt and visually hmm - like I said maybe I setup IA bad

[ioma8]

Those two will likely not give the same results given the same seed and prompt. The random number generation is highly delicate thing.

[Rogal80]

Yes, I agree, but so huge visual difference? And buggy glitches after high-res in IA.

[Rabbitroo0001]

Does InvokeAI support painting an in-painting mask in the web interface or does it rely on external image editors to create the mask?

[illtellyoulater]

you can paint the mask inside the web interface.

[Ladypoly]

I like the UI but there are just many things that don't make it efficient like how Attention/emphasis works. In A1111 you just ave to select text and press crtl + arrow up or down. Same with negative prompts. It's just annoying in Invoke especially on my German keyboard layout to create [ ] for negative prompts if i'm correct.

No option to use other upscaler or models, No automatic saving of all images you create.
And the canvas option lacks a lot of features that infinity canvas and especially workflow and comfort features Hua paint got.

After working with InvokeAI for about 2 hours all I can say is that it looks nice but functionality wise it is really barebone.
The only thing that really excites me is the node based workflow editor the teased.

[Kilvoctu]

In A1111 you just ave to select text and press crtl + arrow up or down.

I had no idea you can even do this, lol. That certainly saves a lot of time (sometimes I don't even do emphasis because can't be bothered to do the typing). Thanks!