Invokeai

The canvas brings layer-based editing to AI generation. You can draw, paint, sketch, and inpaint on separate layers, then move each one on its own. This matters when you're composing a scene in stages instead of accepting whatever a single prompt returns, because you can lock in the parts that already look right. Because every layer stays editable, refining one part of an image never forces you to regenerate the whole thing, which is the part most prompt-only tools get wrong. It's closer to how artists work in Photoshop than to a typical prompt-and-hope generator.

Every generation can be expressed as a visual graph of nodes, so you see exactly how the pipeline runs, which model feeds which sampler, and where the upscaling step sits. You connect models, samplers, and post-processing steps, then save the graph to reuse later. Custom parameters can be exposed as simple UI controls so you don't have to touch the graph just to tweak a value.

InvokeAI supports current foundational models including Flux, SDXL, and SD 1.5 out of the box, so you can pick the model that matches your style instead of being stuck with one. A Model Manager keeps checkpoints, LoRAs, Textual Inversions, and ControlNets organized in one place. That range means you're not locked to one model family when a new release lands.

The software runs entirely on your own hardware, so prompts and finished images stay on your disk with no cloud account required and no third party holding a copy of your work. For anyone worried about handing creative work to a third party, that's the whole point of choosing this tool.

ControlNet support covers depth maps, edge detection, and pose guidance for directing how a generation is laid out. You use these to dictate composition rather than describing it in words. In practice, it's the difference between "a person standing in a doorway" as a prompt and actually placing the figure where you want it.

Batch generation, prompt wildcards, and high-resolution upscaling are all built in, so a single prompt can spin off dozens of variations without a separate app. Wildcards swap in random values from a list so one prompt can produce dozens of variants. Upscaling then takes the winners to print-ready resolution, all inside the same workspace.

The code is Apache 2.0 licensed and developed in the open. Users file bugs, request features, and contribute directly to the roadmap. If a feature is missing, there's an actual path to getting it built rather than a support ticket that goes nowhere. Want a fix faster? Write it.