Ways to use InvokeAI

You can install InvokeAI like any other open-source Stable Diffusion interface, by following the Quick Start instructions on their Github page. After you have it up and running, you'll need to use the Models tab (along the left edge) to download some models, or import models from other folders on your computer.

InvokeAI converts models into a different format than what's used for the other interfaces reviewed here. It'll do this automatically, so you can download whatever SD1.5 based or SDXL based model you want from Civitai.com, and still use it in Invoke. But Invoke can't share a models directory with other interfaces, the way you can with ComfyUI and Automatic1111, for example.

InvokeAI calls the version you run locally its "Community Edition." They also sell access to running InvokeAI on the cloud, so you don't need to use your own graphics card to power the generations.

InvokeAI seems to prioritize interface design more than some Stable Diffusion interfaces. While it's not quite as minimal as Fooocus, it is relatively full featured without being as cluttered as Automatic1111 or Forge.

Looking at the top-level tabs along the left side, there is no separate "img2img" anymore. Image-to-image generation has been folded neatly into a new system called "Control Layers," which is a part of the main "Generation" tab, along with the text-to-image generations. You still can use any image you want for img2img, but it will become a control layer now. The Control Layers section also supports "Regional Guidance Layers," an easy system for regional prompting, where you can paint masks where you want and type a new prompt just for the masked region. While other Stable DIffusion interfaces support some kinds of Regional Prompting, this is a really effective interface for it, letting you easily paint in specifics for certain people within a group shot, or even add negative prompts just to specific parts of your image. It also supports using IPAdapter to provide local image prompts, and using ControlNet to guide generations with sketches or other rough art. See this YouTube video for more about these possibilities.

The Generation tab also has easy checkboxes to generate images that seamlessly tile in the X or Y direction, or both. This option is great for game designers who want to create tiling texture maps. Other kinds of projection textures for specific sets and characters can be created using ControlNet, based on simple renderings or depth maps of the geometry you need to match.

InvokeAI also boasts a "Unified Canvas" that allows you to paint masks and do inpainting and outpainting to your images. You work by dragging around a box that controls the overall area that will be processed at one time, as well as painting masks within that area. It is a little clunky to use, especially with the crudeness of the built-in paint tools. Not having other retouching tools such as a Photoshop-style Remove tool is a limit.  Now that you can do such nice inpainting and outpainting within Photoshop, and the open-source paint program Krita connects with Stable Diffusion via a plug-in, the stakes have been raised somewhat, and these simple masking and painting tools don't seem as impressive as they used to. After generating an area, I would often regret that I couldn't paint back the mask based on what I had generated, but some of these kinds of things don't work the same way they work in Photoshop.

However, once you learn to work with it, the Unified Canvas is a fairly powerful system. By default, it oversamples what it generates, meaning that it computes areas of the image at a higher resolution, letting you add detail to areas of images such as faces and eyes that often need some enhancement. If you want, you can enhance a 256x256 area by regenerating it at 1024x1024 resolution when using an SDXL model. You can easily change which models you are using, with different models contributing to different looks. (You can use models made specifically as inpainting models, but regular SDXL models also seem to work fine for inpainting.) InvokeAI has a slider to fine-tune the Denoising Strength (adjusting how much change it will make to the affected area) and allows you to use ControlNet to produce generations that align with a guide image or a depth map, which could be very useful if you need to make textures that you would align with 3D models.

A tab called Workflows offers a node-based interface, which seems in some ways similar to ComfyUI's interface, although it doesn't seem to have the myriad of custom nodes that make ComfyUI so powerful. InvokeAI doesn't support any kind of animation and its developers don't seem to be in any kind of race to implement all the latest features and technologies. The Workflow editor does have some nice interface features, including a small node map showing the location of all the nodes in the workflow, with a draggable box showing the region of the workflow that is currently displayed on the screen. The Workflow editor also allows you to pull out selected important inputs and put them into a simplified interface for end users to edit, much like what StableSwarmUI allows with Comfy workflows.

The next tab down is Invoke's Model module. This allows you to download and install models from a pre-set list of starter models, or from any URL or local file path. This has been nicely reworked compared to a previous interface, which had been more clunky and DOS-like in an earlier version of Invoke.

Invoke also has a full-screen generation Queue, like a render queue for listing and managing all the tasks that you might have launched or could be waiting for.

In Invoke's Settings, there is a switch to enable the "NSFW Checker." It's off by default, so Invoke will generate nudes or other adult content whenever the prompt calls for it. When turned on, NSFW images are blurred out and have an exclamation point sign superimposed over them. (Most other interfaces seem to have removed this part of Stable Diffusion completely, or in the case of Automatic1111, only support the NSFW Checker as an optional extension that you could install.)

I didn't try InvokeAI's paid service options for this review, but the free Community Edition seems to have a lot of strong points. In some ways InvokeAI seems like a forgotten part of the Stable Diffusion community, but if you want a cleanly designed UI that has a slick interface for regional control and an inpainting canvas as a primary part of the interface, I highly recommend giving InvokeAI a try.