InvokeAI 3.30 for Stable Diffusion (Tutorial)

hi I am Mali and welcome to the channel from a linear workflow similar to automatic 1111 to a node-based workflow similar to comfy and with intuitive features like the unified canvas invoke AI is an all-in-one UI for stable diffusion in today's tutorial I will show you how to set it up and explain the entire user interface first let me say that this is an independent tutorial and I am not saying that invoke AI is better it has its pros and cons whether you should use it or not depends on your workflow it has some good features but it's also lacking in some depending on how you use stable diffusion you may want to use both invoke and automatic 1111 secondly take note of the hardware requirements 6 to 8 GB is recommended for sdxl models it also required 12 GB of system Ram to install invoke go to this GitHub page scroll down to automatic installer and go to the latest release page over here download the installer zip file and extract it before proceeding with anything else ensure you have python 3.10 installed all the links will be in the description while installing python remember remember to tick mark add to path return to your extracted folder and click on this file to enable long path in Windows after that click install.bat to open the command prompt also make sure you have visual C++ libraries installed after meeting all three requirements press any key to continue to select the default location type Y and press enter if you are on an Nvidia GPU go with option [Music] one this takes some time after it is finished choose manual configuration here keep everything on default for the ram setting Reserve about 30 to 50% of your system Ram increase the vram reserve to 1/3 of the capacity accept the license and proceed you can Port models from existing automatic 1111 folders in invoke in starters choose only models you don't have the control net models listed here are non sdxl only in T2i adapters it has only four sdxl models it's missing open pose and my dis depth which I will install manually for the IP adapters I will also have to do a manual install for the sdxl plus and face models Laura and textual inversion embeddings can be manually installed and downloaded as well after you click apply changes it will take some time to download and install all the chosen models once this process finishes close this command prompt and go to the installed folder location here click on invoke dobat file to launch the command prompt type one and enter to start invoke in a browser copy and paste this URL in your browser for the first time it starts with the model manager page go to import models here you can only import diffuser and checkpoint models either via URL or your automatic 1111 folder if you only want specific models go to scan for models and paste in your existing automatic 1111 location don't select the root models folder as you can only import checkpoints now select whatever models you want and quick add them to install Laura vae or embeddings go to the Auto Import folder and download the file in the relevant folder then in the model manager settings click on sync models if the model is compatible it will get added to invoke to install a checkpoint from civit AI rightclick the download link and copy the URL then paste it in the import models for diffusers like the T2i adapter control net model go to the hugging face page if you see this option new use and diffusers it means the model is compatible and will work in invoke copy the green text and paste it in import [Music] models now let's look at the interface on the very left you can see the menu options like the text to image image to image unified canvas workflow model manager and Q the main workflow area is divided into three columns the left column is for prompting and settings the middle column consists of a selected image and some settings if the image is generated using invoke you can use the prompt seed or all settings from the image the right column is your gallery of generated images and assets you may have created or imported in invoke you will find the UI settings at the top right you can add your favorite schedulers and it will show on the top of the list sliders for all options adds a slider next to the input value this option is pretty useful whenever switching between checkpoints invoke will reset the height and the width to the default resolution of the selected checkpoint in The Prompt input you can see a link icon keep this enabled this will concatenate The Prompt and style there is a web page with all the supported Styles and prompts I will show you later on how to use them and invoke they work for thirdparty checkpoints as well and they work very well iterations are how many many times you want the prompt to run if you keep the seed at random it will completely randomize the seed number however if you specify the seed it will generate the images with one increment in the seed value this icon does the exact same function of syncing models from the model manager any image generated from invoke does collect the vae data so if I take that image and put in 1111 I can see the metadata in PNG info invoke has a recall parameter which is used to get the seed prompt and settings from the image the vae data is missing for some reason and there is no way to get the settings with one click for this video I had to switch between different checkpoints and I ended up referring to the metadata too often to check what VA was used with the selected checkpoint the Precision setting is applied automatically based on your GPU the 32-bit floating point is selected for newer cards that support it and 16bit if you have an older [Music] GPU a free aspect ratio allows you to move the height and width slider independently if you choose a fixed ratio it will maintain that ratio perspective of what you [Music] change invoke supports control net IP adapters and T2i adapters however you cannot just download the safe tensor files and get it it working the models have to be diffuser models for each one of them you have to download the model either from hugging face or from their own website typically you put Laura in the positive prompt but there is a separate Laura setting here just click on ADD Laura and you get a Laura strength slider enabling the CPU noise will generate seeds that will give consistent results across different hardware and Os environments disabling it will generate the seed from the GPU Dynamic prompts will divide your single prompt into many prompts I will explain how to use this later on in the tutorial by default all images are assigned to the uncategorized board invoke has excellent Gallery manager ment let's create a new board called cars I can just shift select the images and drag drop them into the new board if you rightclick and delete the board it also gives you the option to delete the images let's check out the unified canvas this is an AI generated image from the previous video by my colleague rightclick and send to unified canvas what I want to do here is make it wide like 16 is to 9 ratio by the way you can drag and hide the left and the right columns with the 16 is to 9 ratio selected I will expand the selection to the full height of the image the compositing settings allow you to fine-tune the Imp painting for this workflow I won't change anything I will explain a bit more in the next workflow the patch match is the best in fill method as per what I tested the scale before processing is used for more Minor Details for such a wide generation It's Better Left To None I am going to leave everything blank except for the negative prompt whenever doing out painting or in painting I prefer more iterations as you can choose and pick the best one let's keep it for five as of now let's hit invoke and if you did not notice I haven't put a seed so the results are not cherry-picked this one seems to have Blended well oh this one is bad there are multiple ways you can do this I just went with the simplest method you can play around with the mask adjustments to make the generation blend better but I wanted to show you what you get without changing any settings you can accept the one you like and keep working on it also you can save the image to your gallery at each step so you don't lose your [Music] progress for this workflow I will generate an image and correct the face we'll show you the Imp painting settings in detail after generating the image you can see an info button at the top when you click that you get the recall parameters I was saying earlier that there is no data for the vae here I don't know if it's a bug or a missing setting because of the recall parameter you can click on use prompt and it will only add the prompt to the workflow it adds all the positive and negative prompts including the style use seed will only add the seed and use all adds everything except VA AE for in painting click on the base and select the mask layer first I will change the ratio to free and adjust the selection to cover the face area this way whenever I generate only the selected area is generated since the selected area is Tiny there won't be any blending issues [Music] over here unmasked will apply the compositing setting to the entire image and The Mask will apply to the mask layer mask Edge will apply only to the edge of the masked area for the face I would be selecting mask Edge for some reason with this checkpoint the Imp painting was low resolution so I will select manual and change the resolution to 1024 what this will do is generate the selected area in that resolution and then downscale [Music] it [Music] we'll stick with the second one moving on let me show you one trick that you can do here and it's very convenient let's choose the open pose control net I can drag and drop the entire image or I can resize the selection area and use this area for the control net and I just got the upper body for the control net whenever you do do this click on the icon that says to set the control image Dimensions to width and height otherwise it messes up the generation I will regenerate this image in an anime style the supported style prompts are on a GitHub page I will leave the link in the description below what you want to do here is copy this and put it in the positive before your prompt the remaining of the prompt copy and paste it in the positive style you should copy and paste the negative text in the negative prompt leave the negative style empty I find that this works the best for sdxl you also want to use control net here to maintain [Music] consistency almost all the given style works except for something too drastic like the isometric one which will obviously not work here let me do one more and show you this one is called a professional 3D model [Music] style I played around with this feature a lot I turned an interior architecture design into a colored sketch pencil drawing really cool invoke has an inbuilt upscaling tool it's located at the top it gives you multiple options and it works great no complaints here to use the dynamic prompts it's best if you decide on the image composure first let's take this sports car for example I can add three colors along with day and night to maintain consistency you should add control net here the dynamic prompt shows six different prompts I will get three colors for the day and three for the night this is very accurate if you do it this way the image composure is very consistent there is a little bleed of the colors and that's primarily because of the prompt style and checkpoint another example where you can use this is to generate the same subject with different fashion outfits invoke also has a workflow editor it is not not exactly as powerful as comfy but it is something the problem is the nodes are mainly open- sourced and contributed by the community I did find a lot of nodes here however it's not the complete package I am not showing the entire workflow for the nodes as it is long and requires another video another thing to note here is that the comfy workflows won't work here the Json has to be invoked compatible I hope you found this tutorial helpful and let me know in the comments if you use invoke AI for your workflow until next [Music] time

Channel: Control+Alt+AI

Views: 4,053

Keywords: invokeai, invokeai install, invoke ai, invoke ai inpainting, invoke ai tutorial, invokeai install windows, invoke ai 3.0, invoke ai outpainting, invoke ai 3.0 canvas tutorial, invoke ai 3.0 overview, invoke ai canvas tutorial, invoke ai models, invokeai walkthrough, invokeai unified canvas, invokeai controlnet, easy installation invoke ai, invoke ai 3, unified canvas, invoke, how to install invoke ai, install invoke ai, invokeai 3, invokeai tutorial, invoke ai install

Id: GZc6gjv4aQ0

Length: 23min 57sec (1437 seconds)

Published: Sat Nov 11 2023