InvokeAI
InvokeAI is the studio-style option among the image generation runtimes on this site, part of local image generation. Where one app gives you a form and another gives you a node graph, InvokeAI gives you a canvas. The unit of work is an image you keep returning to, extending and correcting, rather than a batch of candidates you sift through.
How it works
You generate onto an infinite canvas. A first generation drops an image there, and from that point the tools are spatial: select a region and regenerate it, paint a rough shape and let the model resolve it, or drag the frame past the edge of the existing picture and outpaint to extend the composition. Masking and blending are part of the main interface rather than a separate mode.
Alongside the canvas there is a conventional generation panel for prompt, model, and the usual sampling controls, plus a gallery that keeps every image you have made with the settings that produced it. Models, LoRA adapters and other assets are handled through a manager in the application rather than by dropping files into folders, which keeps the collection tidy at the cost of some directness.
The project also has a workflow editor with nodes for people who want explicit pipelines, but the canvas is the centre of gravity and the reason to choose it. The diffusion machinery underneath is the same as elsewhere.
Hardware and backends
InvokeAI runs on NVIDIA through CUDA, on Apple Silicon through Metal, and on AMD through ROCm on Linux. It ships as a packaged application with an installer that sets up its own environment, so getting a working install tends to involve fewer decisions than the alternatives, whichever backend you are on.
VRAM governs which models load and how large you can work on the canvas. Canvas work is forgiving here, because each operation touches a region rather than the whole picture, so a large final composition does not require a proportionally large card. A 24 GB card covers current models comfortably, and older 24 GB cards remain a good value choice for image work. On Apple Silicon the unified memory pool is shared with everything else running. GPUs for local AI has the comparison.
Fit
Choose InvokeAI when the job is to finish one picture. Compositing, extending, fixing hands and faces, and reworking regions are all first-class here rather than bolted on, and the managed install means less time spent maintaining the tool.
Choose Stable Diffusion WebUI if you mainly want fast iteration on prompts and the widest pool of community extensions. Choose ComfyUI if you need to construct and reuse explicit multi-stage pipelines, or if you want the earliest support for new techniques.
Working with it
A typical session starts with a few generations to find a base image you like, then moves onto the canvas. You extend the frame outwards, mask the parts that are wrong and regenerate them at a lower denoise so the fixes blend, and keep working the same picture until it holds together. The gallery means nothing is lost along the way, so you can back up to an earlier state whenever a direction fails.
Because assets are managed in-app, adding a new checkpoint or adapter is an import step rather than a file copy. Plan for that if you share a model library with another app, and see organising model files for how to keep one copy of everything.