A1111 vs ComfyUI vs InvokeAI vs Fooocus (2026)

If you have a halfway decent gaming GPU and you are not enamoured with sending every prompt to OpenAI or Midjourney, you can run a Stable Diffusion-class image generator on your own machine. The catch is that 'a local Stable Diffusion installer' is not a single thing. It is at least four different open-source projects, all maintained by different people with different opinions about what a good interface looks like.

This guide compares the four serious contenders in 2026 - Stable Diffusion WebUI (Automatic1111), ComfyUI, InvokeAI, and Fooocus - on the things that actually matter when you are deciding which one to install: how steep the learning curve is, what hardware you need, how flexible it is, and what kind of output you are going to get on Tuesday night when you just want a picture for a blog post.

Three reasons people switch from a cloud service to a local install: privacy (your prompts and outputs never leave your machine), cost (zero per-image charges after the GPU is paid for), and freedom (you can use community-fine-tuned models and styles that cloud services won't host).

The trade-offs are honest. You need a graphics card with at least 6 GB of VRAM, ideally 8-12 GB. Initial setup is finicky on Windows and sometimes painful on Mac. And the absolute peak quality of a freshly-trained cloud model (DALL-E 3, the latest Midjourney) is still ahead of what most local checkpoints produce out of the box, although the gap closes every time the community fine-tunes a new SDXL or Flux variant.

For most UK home users with a recent NVIDIA card, this trade is comfortably in favour of going local. Cost dominates - £0 per image vs £8-£30 a month for a cloud generator subscription, indefinitely.

The shortlist criteria were simple. Each tool had to be (1) open-source and free to use, (2) actively maintained in 2026, (3) usable on a single consumer GPU rather than requiring an enterprise A100, and (4) capable of running modern checkpoints including SDXL and Flux variants. That cut the field down from the dozen or so projects you might find on GitHub to the four that the community keeps recommending.

We deliberately did not include the cloud-only generators (Leonardo, Krea, Midjourney) - they are excellent tools but they break the 'local' premise. We also did not include the older Stable Diffusion forks (NMKD, Diffusion Bee, EasyDiffusion) - all still functional, but each has fallen behind on Flux or SDXL support.

Automatic1111 is the project that taught most of the world how to run Stable Diffusion locally. The interface is a tabbed web page hosted on your own machine: prompt, negative prompt, sliders for steps and CFG scale, a model dropdown, and a Generate button. You can be making images thirty seconds after the install finishes.

The reason it dominated the early years is the extension ecosystem. There are extensions for ControlNet (which lets you condition the generation on a pose or depth map), for LoRA loading (so you can apply community fine-tunes), for upscaling, for face restoration, for animation, for nearly anything you can imagine. Some of those extensions are now installed by default; many are still community-maintained.

What you get for that openness is the largest base of tutorials anywhere. Almost every Stable Diffusion YouTube walkthrough older than a year is filmed in WebUI. That alone makes it the right choice for someone who knows they will want to learn more later and does not want to be the one writing the docs.

The trade-off is that the codebase grew organically and shows it. Some features feel bolted on. The default UI is functional but ugly. And it can take a few seconds longer per image than the newer node-based alternatives because the underlying pipeline is less optimised for the bleeding-edge model architectures.

ComfyUI is the other end of the spectrum. Instead of a form, you get a blank node graph. Every step of the image-generation pipeline - load model, encode prompt, sample, decode VAE, save image - is a node. You wire them together yourself. The first time you launch it you will probably stare at it for ten minutes wondering where the prompt box is.

That up-front cost buys you total control. You can build a workflow that does a base generation, runs a face detector, automatically inpaints the face with a different model, upscales with a third pipeline, and writes the result to a watched folder. You can save the workflow as a JSON file and share it. The community has standardised around shared ComfyUI workflows in exactly the way they never did around A1111 prompts.

The other thing ComfyUI does much better is keep up with new model architectures. When a new Flux variant or a new SDXL fine-tune ships, ComfyUI typically supports it within hours. The node-based architecture maps cleanly onto whatever new pipeline the model requires.

The right user for ComfyUI is someone who knows they want a specific repeatable workflow, or someone whose curiosity is bigger than their patience for hand-holding. It is genuinely the most powerful of the four. It is also the only one that can feel intimidating on day one.

InvokeAI sits in a deliberately different space from the others. Instead of a form or a node graph, the centrepiece is a unified canvas with a sidebar of layers, masks, and tools. It treats Stable Diffusion as part of an illustration workflow rather than as a one-shot image generator.

That makes it the best of the four for inpainting (rebuilding part of an image while leaving the rest intact), for outpainting (extending an image beyond its original borders), and for the back-and-forth iteration that a professional illustrator actually does. If you find yourself opening Photoshop to combine multiple generations, InvokeAI is probably what you should have been using.

The trade-off is more memory pressure than the others. InvokeAI runs noticeably more comfortably on an 8 GB card than a 6 GB one, and you will want 12 GB if you plan to keep multiple high-resolution layers in flight. It also has a slightly smaller extension ecosystem - the community has not built every bolt-on the way it has for A1111.

Fooocus was built around an explicit thesis: most users who try Stable Diffusion locally bounce off because the interface is too fiddly, and the actual best practice for prompt-to-image is mostly a fixed set of behind-the-scenes choices. So Fooocus makes those choices for you.

The interface is one prompt box and a Generate button. There are advanced settings if you go looking for them, but most users never touch them. The output is Midjourney-comparable on the default model, and the project takes care of every behind-the-scenes optimisation you would otherwise have to learn (negative prompt insertion, refiner step balance, prompt weighting).

If you have ever talked yourself out of trying local image generation because you do not want to read documentation, Fooocus is the answer. The trade-off is the only one you can predict: if you do later want fine control over the workflow, you outgrow it. The right path for a lot of users is start in Fooocus, move to A1111 or ComfyUI when you hit the wall.

The honest answer is: less than people on Reddit will tell you. A 6 GB NVIDIA card (an RTX 3060 from a few years back, or an RTX 4060) will run any of the four projects on SDXL with sensible step counts. You can generate a 1024x1024 image in 20-40 seconds depending on the card.

The step up from 6 GB to 12 GB is real. Larger context, larger batches, higher resolutions before you have to upscale, room to load LoRAs and ControlNets simultaneously. If you are buying a card specifically for this, an RTX 4070 (12 GB) is the value sweet spot in mid-2026 UK pricing. The RTX 4080 and 4090 give you faster generation but the same headroom.

Apple Silicon Macs work too, particularly recent M3/M4 chips with 16 GB or more of unified memory. They are slower per image than an equivalent NVIDIA setup but they are silent and they idle at almost no power. Installation is fiddlier than on Windows; community guides on running ComfyUI or InvokeAI on Apple Silicon are now reliable.

AMD support is improving in 2026 but is still the most likely path to an evening of wrestling with drivers. If you have a choice of GPU vendor for this use case, the answer is still NVIDIA.

The 2026 honest answer is yes, with caveats.

For the kinds of images most people generate - blog post hero images, art studies, character designs, mock-ups, illustrations of all kinds - a local SDXL or Flux model with a community fine-tune produces work that is genuinely competitive with the major paid cloud services. The community ecosystem of LoRAs and checkpoints means that for almost any specific style or subject you can find a fine-tune that already knows what you mean.

The gap that does still exist is in the absolute top end: prompt adherence on complex multi-element scenes, text rendering, and the polish on photorealistic faces. Newer cloud services (DALL-E 3, Midjourney v7, Imagen 4) have pushed each of these forward in ways that local open-source models have caught up to but not always overtaken.

For most home use, the gap is small and the cost gap is enormous. If you generate ten images a week, the local install pays for itself in months.

Three honest recommendations.

1. Start with Fooocus if you have never run a local image generator before. Two-minute install, no decisions to make, output you would happily ship. Spend a week with it. If you never feel the need for more, you have your answer.
2. Move to Automatic1111 when Fooocus stops feeling flexible enough. The transition is straightforward; most of the community walkthroughs you will want to follow are filmed in A1111. The extension ecosystem will keep you busy for a year.
3. Reach for ComfyUI when you want a workflow you can save, share, or fine-tune. The investment in learning the node graph pays back the day you realise you need to repeat the same five-step pipeline twenty times.

InvokeAI sits slightly outside that path. If your goal is illustration or art - not just isolated images - it is the right starting place, full stop. It is not better or worse than the others; it is built for a different job.