InvokeAI - Workflow Fundamentals - Creating with Generative AI

[@zeropoint216]

Thanks for these. Invoke is such an incredible platform

[@robertmichaelgoodwin7134]

Please keep these coming! This and the last Fundamentals video have been especially wonderful. The balance between theory and practice is admirable. Bravo!

[@PieterLaroy]

Great and educational once again :-)

[@Xeraser2]

I think people would really benefit from a simple tutorial on how to use the unified canvas to create seamless pieces from scratch or seamlessly outpaint pre-existing generations. It's the main appeal of Invoke for a lot of people but in the canvas drivethroughs like #2 it's kinda just assumed that the viewer already knows how to do it and do it optimally.

[@TimothyMusson]

Thanks for these excellent tutorials - I'm loving all the new features in InvokeAI! :)

[@MarkJoel1960]

This just keeps getting better and better!

[@dschony]

Thanks for your work! ⭐⭐⭐⭐⭐ And it's much faster than in the editors, even with higher resolution!

[@DanielSchweinert]

Wow, my respect! A fantastic explanation about the workflow. Time to really look at these nodes. Thank you very much!

[@hamzasidki1090]

amazing explaination, invoke is one of the best if not the best ai platform

[@ViralKiller]

didnt understand a word but looks amazing

[@florian7347]

your videos are great! thank you and keep going! 🍻

[@Joshua-rq3om]

I'm not sure if this is possible, but it would be helpful if the node workflows could autogenerate for those using the standard UI first. This might make it easier to understand, initially, for the less technical folks. Keep up the great work!

[@Adreitz7]

20:40 Pro tip regarding this error: since the dimensions of the noise and the first stage latents must always be the same in order for Denoise Latents to function, you should connect the output Width and Height from Resize Latents to the Width and Height inputs of the second Noise node. It should also be pointed out that Resize Latents is prone to doing things that the model doesn't like (particularly with SDXL) and garbling the output in unpredictable ways. It is generally better to output the first stage image, connect it to a Resize Image node to upscale it, and then to an Image to Latents node in preparation for the second stage.

[@16bitPRNTZ]

Loved this vid, understanding workflow is extremely helpful (I was in-between GPU hardware upgrade when comfy was out so I hopped back in and never completely understood work flow)

[@mada_faka]

wow invokeai getting improve, i hope you guys can match with midjourney

[@0x44Monad]

Your presentation by 1:42 reminds me of the idea of a universal translator. Given an intelligence comes across a "model" would one be able to infer enough information within that latent space that is embedded to transpose it to any language?

[@0x44Monad]

Does a latent space represent the 'minimally compressed coordinate space' that any intelligence can leverage as the data layer then translate it across any domain from a projection that makes sense of that information? In that sense then its kind of arbitrary as any meaning can be imposed on the symbols but there would definitely be a natural order or structure to them based on the groupings/geometries/shape.

[@GaelLendoiro]

Nice!! keep going. Thanks!

[@RobertWhite-q8x]

This is awesome. What would the process look like for outpainting?

[@ShinyVeggie]

Just a random thought. It would be cool to have some sort of standardised spec so it would be easier for extension/plugin developers to create the same plugin for multiple uis. Might not be feasible and too limiting but would be great if for example it was easier for researchers to bring new features to all the uis. Auto1111 feels like a stretch but cross compatible nodes for invoke and comfyui feels possible.

[@ashleylm]

I'm one of those "non-technical people." It would be helpful for me if you explained what the Workflow is and does. I understand the text to image tab. I get image to image. Unified Canvas makes sense. I'm at ease with the Model Manager. But this Workflow seems like a Really Really Complicated Way to do the stuff that InvokeAI already so brilliantly designed its user interface to do. So what is Workflow and why would I want it? I honestly have no idea, and I've watched the video twice, poor non-technical dullard that I am! (You touch on this for about 15 seconds at around the 7:30 mark, but don't give examples of why we'd want to use workflow other than to do something beside Text to Image for which we may as well use your first tab. But what is the something? I can't conceptualize it). Sorry! And I love your program so very much, and want to master it completely.

[@quincy3d69]

it would be super nice to be able to resize and rearrange the fields in the left linear view . some time i just forget to add one parameter in my workflow and i have to add all the fields once again. if i could just add it later and drag it where i want would save a lot of time. and please some reroute thingy to control the node noodle mess

[@bgtubber]

Even at 1080p the video looks like 720p and it's hard to make out text and other graphics in the interface. ☹

[@vintagegenious]

You need to do more node videos with that form

[@ai-jm6yq]

is this compatible with comfy workflow?

[@vincentcarlucci1259]

Node based systems are extremely difficult to work with, and I am rather lazy, so my question is: is the output from this node system in any way better or even significantly different from what can be achieved with regular invoke?

[@Adreitz7]

This video shows elementary cases which, in theory, would output the same results as the linear workflow with the same settings. If that's all you want to do, you don't need to use it. The benefit is the ability to tinker and experiment, creating multi-stage workflows possibly including multiple models, controlnets, etc. The only requirement is that the whole thing be mathematically sound (which the configuration error at 20:40 was not). One setting which is not available in the linear workflow, though, is tiled VAE encoding/decoding, which can significantly reduce peak memory use with a slight performance penalty.

[@vincentcarlucci1259]

The important word in my comment was 'significant'. Certainly you can tinker a lot more but would the resulting images be significantly better or different? I don't see how they could be.

[@nomorejustice]

Unfortunately for low-spec computer users like me, it is very difficult to access Invoke AI. I am willing to subscribe if it is run on the website hopefully it can and I hope it happens

[@mongini1]

If only invoke would work for me... No matter which python version I use as primary, I don't get it to start. just tried a fresh install of invoke, but everytime i try to launch it i get: AttributeError: type object 'GraphInvocationOutput' has no attribute '__annotations__' i got Pythong 3.11 installed, ran the installer as admin to avoid issues, installer finished sucessfully but on every try to launch invoke this will pop up

[@daffatahta9059]

You could try using python 3.10. Why python 3.10? Since python 3.11 is still new a lot of modules isn't updated yet and that version difference probably causing the error on your part.

[@daffatahta9059]

If you have multiple python version installed you could re-arrange the python 3.10 on top of the python 3.11 in Path on your system

[@mongini1]

@@daffatahta9059 already tried that too... tried with 9,10 and 11, always the same

[@invokeai]

Join our discord at discord.gg/ZmtBAhwWhy and we can help :)

[@mongini1]

@@invokeai I posted that in the help section 6 days ago on discord