InvokeAI 3.0 - NOW with Nodes, Controlnet and Lora Support
Vložit
- čas přidán 20. 07. 2023
- InvokeAI 3 is out. Here is why this might be you favorite new UI. InvokeAI 3 added a lot of new fuctions and optimized their UI. Especially the addition of Gallery folders, Loras and ControlNet really helps. InvokeAI now also has NODES. The Nodes are not just perfect for every advanced user. You can also create your own nodes with a bit of code and the help of ChatGPT
#### Links from the Video####
Python 3.10.9 www.python.org/downloads/rele...
InvokeAI installation: invoke-ai.github.io/InvokeAI/...
Microsoft Visual C: learn.microsoft.com/en-US/cpp...
InvokeAi latest Release: github.com/invoke-ai/InvokeAI...
#### Join and Support me ####
Buy me a Coffee: www.buymeacoffee.com/oliviotu...
Join my Facebook Group: / theairevolution
Joint my Discord Group: / discord - Jak na to + styl
I may give it a try one day. But sometimes I feel like people are moving from thing to thing too fast to fully get the most out of the previous tools they were using.
This is the problem with A.I. People moving too fast, not knowing fully how to use the tools properly.
You can use stability matrix to have multiple ui's
sure, but that is because it's new and everyone still tries to find the best way to use it
I’d also argue that we don’t know what the best UI for AI looks like yet. As things evolve and features survive or die, I’m sure interfaces will standardize and it’ll be easier to master what we have. It’ll just take time.
@@SG-jm7np its not moving too fast, its just that content creator need to create content and so do it for stuff not that important. While you guys think you have to use everything they talk about.
I hadn't realized you can drag assets in to the ControlNet reference input! I'd been downloading them and re-uploading. 🙈 Thanks for the demo!
Love your videos, it helped me alot. Big thank you!!!
I really like the Invoke interface. I have recently migrated to Comfy, but I might have to reinstall Invoke and play around with this.
this is great, I love using invokeAI I would love it if you do a video going more indepth with controlnet and upscaling techniques with invoke and the different controlnets
This UI looks rather nice. Thanx
Thank you very much for sharing. It would be even more convenient if you included chapters in your videos so viewers can go directly to the part they want. For example, I could have saved the first 10 minutes to quickly check out what the new features of InvokeAI look like.
100%!!!! I was thinking the same thing...takes the content creator less than a minute to include these.
What a beautiful interface.
Once I am used to comfy ui I don't want to go back to traditional interface anymore. Node base workflow does take time to setup, but it's very flexible.
I agree, this one is much too slow for my tastes. I havent mastered training on ComfyUI yet. Nor have I found a good page that explains how to do it.
This is great work!
I highly recommend InvokeAI.
I was an absolute beginner when I installed it, and with help from the people in their discord server I've now generated over 30k images, of which I've judged 2k to be flawless. That's a 15% success rate, and I'm rapidly getting better at making the right prompts and combining the right LoRA's. Definitely worth checking out!
Jesus Christ lol. What type of images are you generating, and what are you doing with them?
Great breakdown. Well done.
thank you much lol
Mac Users (on M1 or newer Macs):
- Precision: float32
- Disable the "Enable xformers support" - it's for NVIDIA chips
Love it! Going to install it
Life is so much simpler with Stability Matrix. I don't understand why it's not more popular amongst SD users.
Ironically, the most advanced and intelligent software applications of the present exhibit an outdated, non-graphical interface in their installers, making the installation process quite challenging for ordinary users.
I really like the Invoke interface.
The outpainting system from invokeAI is symply the best right now
@OlivioSarikas interesting video, but I have few questions: I wanted to move to Comfy UI, but it doesn't have "Batch size" (not "Batch count", as "Batch size 2x faster") function and also missing "Only masked" in Inpaint function (it generates whole image instead of only masked area). What bout InvokeAI ? From interface i didn't see those functions...
Hey great summary, really helped with the models portion, but how do we add more loras, vae's and controlnets? not entirely sure where to put those files. thank you
Olivio, I have a special wish (I know its not yet Christmas time): I would love to see a video about the various SD models, checkpoints and LORAs compared. So much has been going on and many checkpoints come and go. I understand that there are many videos featuring individual models, but it would be nice to see a recap and an overview of what is out there now for the various SD model versions. Especially which model - checkpoint - embedding - LORA or combo to use for various use-cases like landscape photography, portraits, illustrations, tiles and patterns aso. Maybe geared towards professional designers (like it used to be) and less towards the anime crowd?
+1
+1
a lot of work, but but would be nice to see a huge comparison. There are some smaller comparison videos out there (with ratings on hand quality etc)
Very useful!
Pretty brilliant tool
Nice tutorial!!
Great video.
Great once again Olivio. Is there a icon that should appear on the taskbar after the instal, as in the earlier version? Thanks.
Thanks!
I much prefer Invoke as opposed to A1111 just for the UI. It does lag far behind in terms of extensions and features, but 3.0.1 is great with nodes and controlnet. I think it works great and I'll stick with it because it's easier to work with for sure. highly recommended for anyone that is starting to get in to AI generation now
Same here. I actually started with Invoke before I even knew A1111 existed... I tried A1111, but the UI and all the settings are just overwhelming. I much prefer Invokes clean and organized interface, even though I miss out on some features, but I already got a nice workflow with Invoke now, that delivers amazing pictures. Only problem for me is, everytime I updated Invoke, everything broke.
Invoke was extremely buggy and selection button freezing all the time last time I checked a few months back.
@@dadabranding3537only ever time i tried invoke it broke and nothing worked with it so i gave up lol
Thanks, this looks really cool. Can you also make a link to all the models you're using in eg automatic1111 already, meaning not only the checkpoint models, but also loras, vae files, control net models, maybe upscalers as well? Would be cool to only have one centralized source models folder where every UI has access to so that you do not need to download and maintain anything double. Thank you.
Thanks again for your work and excellent presentation. I installed it on my m2 Mac with no problems. Another 30gb download brings me up to 879 GB for July.
I tested the speed of InvokeAI against an app called DiffusionBee on the Mac, and it was a tiny bit slower. That's not bad, and DiffusionBee is quite well-optimised for use on a Mac.
But it is like Automatic1111 & Vlad can process the images in parallel. So far as I can see, this UI cannot. I assume it will be just a matter of time until they add it.
Isnt that built in by default in SD? I haven't downloaded and tested this program yet but i find it hard to believe they missed that. Maybe its somewhere in options when you create your node.
great , thank you
You are welcome!
does this have other extension supports like deforum or infinite zoom?
Not at the moment, but probably with nodes
Invoke can convert your models to diffusers, which has the added advantage of loading them very quickly later. This might not be ideal if you are using other interfaces, as it will more than double the space on your hard drive.
Gitgud is now the normbness of what artist needs to get thru to make a living ect lol
I loved the interface. But I'm having stability problems in inpaint and I haven't found the Scripts, which are very useful in the A1111, mainly the SDupscale.
Anyway you can do a video on how to remove/delete metadata from an image. Thanks. Great video once again!
nice video but did I miss the section on how to get control net to work like where do you download it from and also the loras do they go into th sd-1, sd-2, sdxl or sdxl-reiner folder?
I see 20 Images and 100 Steps maximum on the sliders in some videos for Invoke AI 3.0, are they limiting the steps now or planning to?
And upscaling is limited to 1536max, right? Considering updating or not, thanks!
What I want to know is: have the same bug of automatic1111 and forks when you change the model and RAM still increasing until crash? Mainly when changing to XL and come back to 1.5
I'd love to see a head to head comparison of the image outputs between the most popular Stable Diffusion models. With so many options now, it would be great to know which ones produce the best images.
Really depends on the subject matter, prompts... there's too many variables to do a fair comparison
Plus, what do you mean by "best"? What's better, stylized images or realistic ones? 🤔
It's not about the image output, it's about the workflow. What do you need and how fast do you need it.
Best is relative.
But you can do x/y/z plots with different models to see what works for your prompts ad what you are trying to create.
@@OlivioSarikas I disagree. For me it is all about the image quality. If it takes 20 minutes to make the perfect image in one ui and 2 minutes to make 10 "meh" images in the other ui, I go for the first one. The 10 "meh" ones would end up in the bin anyway. Quality over quantity all day every day.
Are there any workflows for creating a 3d character mesh for a video game with stable diffusion or any other AI?
Thanks for this video. I really liked invoke but eventually switched over to automatic 1111. Haven't tried comfy (but I haven't seen that it has CPU support - I have an older Radeon card without Rocm)
comfy does have cpu
Ugh, how long does it take to process single 768x512px on CPU? Is it even worth it?
So this is basically a web ui, a workflow thing, you can still use all the same underlying models, loras, textual inversions, etc, like in, say, A1111? Is there anything that sets it apart if you consider that you already have ControlNet extensions and ComfyUI in A1111?
there is no option for an upscaler in the img2img right? So HighresFix on Auto seems to be better.
Great video, but it looks like you missed the ability to convert a ckpt or safetensor model to diffusers. You can merge any model you want after converting it to a diffuser, and it's a very simple procedure. In the model manager, pick a ckpt, and in the upper right is a convert model to diffuser button. Hit that and it works its magic. I've merged so many models (photon was the last to get added) I can't even keep track anymore...
ah yes, i forgot about that. good point!
I want to use SDP not Xformers. Do I accept Xformers during the install and change that later in the UI, or just leave XFORMERS disabled during set up?
OK probably a n00b issue but I cant seem to progress past the invoke startup options. I have everything down to "folder to recursivley scan for new checkpoints..." but no input for entry :( i can close it and run the invokAI.bat but worried Ive missed a load of settings. UPDATE: So i think I got it working but is it supposed to take a long time? I have a 3090 founders edition and have been finding faster outputs with A1111
Olivio. what is the key to go to the next page in the cmd? I often struggle with that because it's never mentioned ;;;
@5:17 - there is no button below the choices for APPLY (at least in the Windows installer, using Windows 11), there is only 'Apply Changes & Restart'
*Edit, finished it, it kept window opened and installed after it closed, so no issue
What does it mean "A matching Triton is not available, some optimizations will not be enabled." when it says that for Invoke 3.0? I keep getting that warning, but I have no idea how to fix it.
Thank you for sharing. How to link Lora model folder? How to link VAE model folder?
Does it work with the Intel Arc these days? Wish more AI tools would support the Arc Series as its just a cheap way to get into AI. Like what other GPU is there for 350 bucks with that performance and 16 GB Vram?
currently using Auto1111, should I switch to comfyUI, Invoke using comfyUI or just download the Auto1111 ComfyUI extension? (for better control)
oh nooo ive been out of the game for a couple of months and WHAT IS INVOKE. Its pretty cool! I have an extension that grabs from civai links in 1111 but dang, now im TORN
Folder to recursively scan for new checkpoints, ControlNets, LoRAs and TI models ( autocompletes, ctrl-N advances): ???Why doesn't it show anything at the bottom? Help！！
Is this portable as well? So I can just move the files after installation to lets say a external hard drive?
Btw, I'm getting a long error starting with "ERROR: HTTP error 403 while getting https"pytorchURL".
Could you make a tutorial on how to install and use a High Res upscaler for InvokeAI?
I shall wait for colab version
Is it possible to compare models via x/y/z prompts like I can on automatic1111?
for text2img A1111 is the king. for inpainting and outpainting InvokeAI is the goat. I got both of em running and use it for this very purpose.
Hi great video, I’m a little new to this and don’t have a lot of money to spend on a you, any recommendations for something in the $100 range that will work well?
Can you tell how you got past "A matching Triton is not available, some optimizations will not be enabled. Error caught was: No module named 'triton'. This is where everything ends for me. It tried to fix itself but to no end. The process keeps repeating. I have read a lot of blogs and posts and I understand that a lot of people have this same problem.
During installation, there is no apply button in the middle, so I had to click on Apply Changes and Exit...
I loved INVOKE mostly because of the outpainting / inpainting feature. I have not found anything alike in Automatic1111 that works so smooth and does what it should. But having two systems on my PC (SD/A1111 and Invoke) and maintaining all those Loras, Checkpoints (which take quite some space on HDD) made me switch over to A1111 again.
If there's a chance to use the SD checkpoints and link to that folder, that would be highly appreciated. Converting / duplicating them for Invoke really drove me nuts.
there smt called symbolic link that'll solve all your problems, Just move all your files Checkpoints, loras etc to an organised folder on your pc or external drive, then you can do the symbolic link to A111 and invoke AI those take less than 100kb combined
For comfyUI you can specify folder where models are, simply input path of a1111 models folder.
For programs that dont let you choose model directory you have to use "symbolic links" as mentioned in post above.
You can use stability matrix for it, it is quite convenient solution for many sd UI`s
Nodes are cool and everything, but, any chance to get to use SdXL in A1111 someday ?
thanks for your great tutorial for invokeAi, I am newbie on sd but its too many items I really don't get it, and the main problem is its no good for AMD card... i had tried oliver version but I don't how to convert the savetensor and ckpt files to onnx ... , the InvokeAi seems user friendly , I would try it in coming days
I have tried with Invoke 3.0 many times to move the .cache. folder from my C drive to my D drive. I found a place to change the path, and I moved my cache to the D drive. But when I ran Invoke again, all I got was rebuilding the .cache. on my C drive again! I have tried everything I can think of, but found no answers on the internet either. I was hoping you could help, maybe?
This appears to be much better than Automatic 1111. And I understand that it uses less VRAM.
How can I create 2560x1440 Images?
One thing I don't like is that it doesn't remember settings per tab, if I set a size in text to image, then use image to image with a different size it will change the text to image dimensions as well
You're right. This is incredible. The node system is nuts! The Shared RAM and CPU features is great and generates images way quicker than Stable Diffusion.
Invoke AI is StableDiffusion
Great video! I have a question: after the installation I can no longer see the previous images I had generated. Is there any way to have them show up on the gallery again?
This feature is missing... 'A migration path to import 2.3-generated images into the 3.0 image gallery'. I used Automatic 1111 gui/png info to extract the data then cut and pasted into InvokeAI. Bit clunky. But its a workaround.
Although i have python installed my system cant find it. (the path is in enviromental variables) how can i fix it? i cant generate anything ... i cant even install it.
I have an RX 6800 AMD GPU which AI tools will support it? it has 16GB VRAM
Thanks a lot! Is there any way to add controlnets after the installation? I accidentally clicked on agree and exit and cannot add the other controlnets anymore
Yes, when you start invoke it gives you the option to download additional model. Chose that and select all the controlnet models and select apply
cool ui, no build-in tiled vae decoder is a deal broker to me (unless I haven't figured out how it works). Increase the image size a bit, everything just falling apart.
following the instructions the install stops at page 1 of 2. It doesn't show the Autoimprt Folder or the LICENSE info. Is there something I missed? Thanks
for me install stopped at 3:04 and then blue screen of death appeared
Fullscreen the terminal and then restart the installation, that solved it for me.
ImportError: DLL load failed while importing cv2: The specified module could not be found.
this is the error i keep having. Same as stable diffusion, somethings wrong with the python. Tried all the possible resolution for cv2 error but still the same.
please help me with this error.
Olivio, how does the Cabby model work for you, since mine always shows the resulting image as transparent.
Nevermind, turns out you shouldn't use INPAINTING models. :)
InvokeAi is great in some features, but as i see it does not have a weight control hotkeys like 1111...
I've started my AI journey thanks to invoke, but switched to automatic1111 to learn to use controlnet. Now that invoke 3.0 has implemented controlnet (and much more) I wanted to go back, but since I've been confident to use automatic1111 I wanted to use both. but I see that they use different versions of python to work correctly. what should I do to use them both? any suggestion?
You might already know this as it's been four days, but you can get standalone versions of both Invoke AI 3 and A1111. These should install completely separate versions of each with all their dependencies. I don't have Invoke yet (downloading it now) but I have standalones for A1111 and ComfyUI. I'd post a link but YT doesn't like them so if you still need them, do a googly for them with 'standalone' added. 🙂
I can use Lycoris safetensors in invokeAI?
I started my SD journey with invoke, i love the interface, but they are too slow to catch A1111. I am so happy that i finally switched to A1111, although the interface is lame, but the dev community is so huge, all the new technologies come up all the time. Meanwhile, i had been begging invoke to support lora and control net for so long until now.
exactly that!
Yeah this was me too.
Automatic 1111 still doesn't have the unified canvas though.
What if even after following all steps still cmd tells me i dont installed my python?
InvokeAI had the 4x upscaler and 2 face-restore options, which worked really well together. Where are they? I loved InvokeAI for having a highres 2048x2048 picture with perfect restored faces in a few seconds. If they cut this out, it is a massive downgrade to the workflow.
can you do a video on automatic 1111 vs comfy vs invoke ? please
Trying to install it but i keep getting "[2023-07-22 13:42:41,404]::[InvokeAI]::ERROR --> Diffusers model stabilityai/stable-diffusion-2-1 could not be downloaded. Skipping.". Any idea whats going on? thank you for the great content!
Unfortunately InvokeAI has no api access to the server which is a shame. Would be nice to integrate it with other tools calling into it.
It's my first time using InvokeAI. What I have noticed so far is, that InvokeAi seems to generate much more photo realistic and detailed images than AUTOMATIC1111, even without any words like masterpiece, high quality or something like that, only with positive and negative prompts. But I don't know why, because both use SD at the end?! Is there any processing in the background? Images are a lot sharper and more realistic. Faces are also much better, even in full body shots. (But it seems like they have removed the option to restore them individually). It's also faster and more memory efficient.
yes, for some reason Invoke can often generate better results with the same settings and prompt
same thing for NMKD, all UI based on invoke gives much more better results ( that will be never achieved by Automatic1111 even if you use the best upscaler there) and the processing i s very fast compared to Automatic1111. now that invokeAi supports Lora and Controlnet i think i will never use Automatic1111 ( maybe i will just use it as a toolkit)
NODES!
dynamic prompt is not a different name for the xy but its an extension that is already available in 1111
I didn't say it's like XYZ plot, i said it's like search and replace in the XYZ plot
#### Links from the Video####
Python 3.10.9 www.python.org/downloads/release/python-3109/
InvokeAI installation: invoke-ai.github.io/InvokeAI/installation/
Microsoft Visual C: learn.microsoft.com/en-US/cpp/windows/latest-supported-vc-redist?view=msvc-170
InvokeAi latest Release: github.com/invoke-ai/InvokeAI/releases/tag/v3.0.0
#### Join and Support me ####
Buy me a Coffee: www.buymeacoffee.com/oliviotutorials
Join my Facebook Group: facebook.com/groups/theairevolution
Joint my Discord Group: discord.gg/XKAk7GUzAW
I think you missed the Discord link for InvokeAI. Great video though.
I got this error (A matching Triton is not available, some optimizations will not be enabled. Error caught was: No module named 'triton')
You can ignore that.
How do I disable Image autosave on Invoke ai?
Hey Olivio, you have a great channel. Appreciate your content. Question: How can one learn about all the stuff like Loras and all the fine-tuning stuff, and how it all comes together so you can actually start using node-based workflows effectively. I have seen quite a few videos where people show you the interface, but not very much that insofar as teaching you what all the various components ARE and how they can be mixed and matched. I find if you go beyond the basics, it's very confusing. Are you aware of any resources that goes into everything having to do with models, loras, fine-tuning, controlnet and so on, so that a person can go from basic to more advance SD user? Thanks in advance.
by watching my channel. i have shown many model, loras and embbedings in the past
@@OlivioSarikas Thanks. I will look for the videos. Basically it would be awesome to have a primer video that explains all the concepts from A - Z.
Does anyone know - I want to try this out, but I used Python 10.3.6 to install Automatic1111, if I follow this guide will this mess up my ability to use Automatic1111 as it uses different Python 10.3.9 version ?
you can use anaconda/miniconda and create separate python environments for whatever programs you want.
I would like to know this too please :)
@@schmutly me too
Yeah better create separate python enviromments as even if a1111 will be fine the clusterfck of dependencies and addons we usually install into a1111 may not be.
After some research, here is how(with miniconda):
Use conda create(in cmd) to make a new environment:
conda create -n py310 python=3.10
(This creates an environment called "py310" with Python 3.10)
Use conda activate py310 to switch to an environment:
conda activate py310
To switch back to the default base environment, you can use:
conda activate base
If -for any reason- no longer need this environment, use:
conda env remove -n py310
hello I don't have this paragraph in cmd panel: I accept the CreativeML...
DID YOU BUY THAT SHIRT AT DAN FLASHES???
is there any way to generate images with a hi res fix like automatic111? or do you need to just use img2img?
it seems like there is no highres fix. but i'm sure they had that in the past.
@@OlivioSarikas hmm they may put it back eventually
@@OlivioSarikas yeah, it was in the previous 2.3 version but it was removed in streamlining things for 3.0. Might come back as an extension later, but it sounds like people have been using ControlNet to solve the composition problems that hires fix was trying to address.
hey mate, i keep getting an OS error. no one seems to know how to fix it on windows
Simpler = better. No proper UX was made on the AI tool so far. Some web based services try to address UX pains (Leonardo or tensor art) but in general experience in AI clients needs a lot of love. Invoke go in the right direction.