InvokeAI

Last reviewed Aug 26, 2026 · 794 words

Pick InvokeAI when the job is fixing an image rather than generating one. Its unified canvas, where you inpaint, outpaint, and layer control images in a single editor, is the best of its kind in open source, and nothing else in the self-hosted image generation field matches it for iterative work. If you want to run a batch of 200 prompts overnight or wire together an unusual pipeline, ComfyUI is the better tool and InvokeAI's own node editor is a distant second.

VRAM is the budget, and 8 GB is the floor for SDXL

The catalogue lists 8 GB of RAM as the minimum, but the number that actually decides your experience is GPU memory. Rough working figures, from my own cards rather than a benchmark: Stable Diffusion 1.5 models are comfortable at 4 GB VRAM, SDXL wants 8 GB to avoid constant offloading, and FLUX-class models need 12 GB or more even with quantised weights. Below those lines InvokeAI still runs, but a 1024-pixel SDXL generation crawls from about 10 seconds to over a minute as layers shuffle between GPU and system RAM. The VRAM maths post covers how to estimate this for any model.

NVIDIA is the well-trodden path. AMD works on Linux through ROCm, and Apple Silicon works through MPS at roughly half the speed of a comparable NVIDIA card. Windows with AMD is where I stop recommending it.

Two installs: the launcher for desktops, Docker for servers

InvokeAI ships a desktop launcher that installs Python, the right PyTorch build for your GPU, and the app into a directory you choose. For a workstation that is the correct choice; it handles the CUDA version matching that ruins most manual installs. For a headless box, the Docker image is cleaner:

services:
  invokeai:
    image: ghcr.io/invoke-ai/invokeai:latest
    ports:
      - "9090:9090"
    volumes:
      - ./invokeai:/invokeai
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    restart: unless-stopped

The web UI comes up on port 9090 and everything persistent lives under that one root directory: models/, outputs/, and a SQLite database of boards and generation metadata. Back up the root and you have backed up the installation. Expect the image and a first model to cost 15 to 20 GB of disk before you have generated anything.

The canvas is the reason to choose it

Most Stable Diffusion front ends treat inpainting as a mode you switch into. InvokeAI treats the canvas as the home screen. You place a generation, mask a region, regenerate just that region with a different prompt, extend the border to outpaint, drop a control layer (depth, pose, edges) on top of a section, and the history of every step sits in a gallery beside it. Compositing a product shot or fixing hands on a portrait takes minutes rather than a cycle of export, edit, re-import. If you have used Photoshop's generative fill, the workflow will feel familiar, and it runs on your own hardware with any model you choose, which is the honest case for it over Midjourney.

Models: the manager is good, the disk is not infinite

The model manager imports from Hugging Face and civitai by URL, converts formats, and tracks which base architecture each checkpoint belongs to so the UI hides incompatible LoRAs and control models. It is the most polished model management in any of these tools. The cost is that each SDXL checkpoint is around 6.5 GB and each FLUX checkpoint over 20 GB, so a hobbyist collection of a dozen models plus LoRAs reaches 100 GB quickly. Put models/ on a separate volume from day one.

Boards organise, they do not isolate

The catalogue tags this as multi-user, and it has boards for grouping generations and a login. Do not read that as multi-tenant. Every user shares the same models, the same queue, and the same GPU, and there is no per-user quota. It suits a household or a small studio where everyone is trusted; it does not suit offering generation to strangers. The comparison in Automatic1111 vs ComfyUI applies here too: none of the three were built to be shared services.

What I'd do

A used RTX 3090 or 4090 for the 24 GB, the Docker image on a Linux box with models/ on its own disk, and InvokeAI as the tool for anything involving a canvas. ComfyUI alongside it, sharing the same model directory through a bind mount, for batch runs and experiments. Apache-2.0 licensing means nothing in that setup is time-limited or phones home, and 28,000 stars of active development mean new model families arrive within weeks of release.