Awesome Open Source AI
Editorial ranking · 10 toolsResearched Sep 2026Non-sponsored · Open source

Top 10 self-hosted image generation tools

This list is for people who want to generate images on their own GPU with open models such as SDXL, FLUX, SD 3.5, Z-Image and Qwen-Image, through an app or UI rather than a hosted service. The picks are the interfaces and runtimes; models come up where they decide what a tool can run. The code is open source, but many popular model weights are not, so check the weight licence before you sell the output.

The #01–#10 order is our editorial recommendation for this audience, not a benchmark or a popularity ranking. Current model support, active maintenance in 2026 and hardware reach drove the order, which is why AUTOMATIC1111 and Fooocus are not on it. If the same GPU also runs a chat model, see the self-hosted ChatGPT alternatives guide.

Flat illustration of a llama in a beret and scarf painting at an easel, where the canvas shows an abstract diagram of connected nodes and wires

How we picked and ranked

  • Open source and local: Every pick has an OSI-licensed codebase and runs on your own machine. Easy Diffusion was left out because its licence adds OpenRAIL-M use restrictions, and Draw Things because its app is closed source.
  • Maintained in 2026: Each pick had commits or a release within the last few weeks of September 2026. AUTOMATIC1111 (master frozen since July 2024), lllyasviel's Forge (last commit June 2025) and Fooocus (bug-fix LTS, SDXL only) failed this test despite their star counts.
  • Current model support: We ranked higher the tools that run 2025–2026 families such as FLUX.2, Qwen-Image, Z-Image and Krea 2, not only SD 1.5 and SDXL.
  • Hardware reach: Documented support for AMD, Intel and Apple GPUs and for low-VRAM setups counted in a tool's favour. Figures come from each project's docs; we did not benchmark.

At a glance: the 10 picks

Each card links to its review. Full generation UIs are #01–#04, workflow-specific tools #05–#07, and runtimes, code and Mac apps #08–#10.

ComfyUI logo

ComfyUI

Node workflow engine

A node-graph engine and API that usually supports a new open model family before any other UI, and is the backend several tools on this list run on.

GPL-3.0New model families first
SwarmUI logo

SwarmUI

Generate-tab UI

A prompt-and-parameters UI that drives a bundled ComfyUI backend, so you get ComfyUI's model support without building graphs, plus a raw Comfy tab when you need it.

MITComfyUI backend, form UI
InvokeAI logo

InvokeAI

Canvas studio

A polished local studio with a layered canvas for inpainting and outpainting, a gallery, and node workflows, under the most permissive code licence of the full UIs.

Apache-2.0Layered canvas + workflows
SD.Next logo

SD.Next

All-in-one web UI

A Diffusers-based web UI with the longest list of compute backends on this page and its own SDNQ quantisation for fitting large models into less VRAM.

Apache-2.0Widest GPU backend list
Krita AI Diffusion logo

Krita AI Diffusion

Painting plugin

A Krita plugin that generates into selections and layers of a real painting app, with ComfyUI running locally as the engine.

GPL-3.0Generation inside Krita
Forge Neo logo

Forge Neo

A1111-style web UI

The maintained continuation of Stable Diffusion WebUI Forge: the familiar AUTOMATIC1111 txt2img and img2img tabs, updated for FLUX.2 Klein, Z-Image, Qwen-Image and Krea 2.

AGPL-3.0A1111 layout, 2026 models
Stability Matrix logo

Stability Matrix

Package manager

A desktop launcher that installs and updates ComfyUI, SwarmUI, InvokeAI, SD.Next, Forge variants and training tools with one shared model library.

AGPL-3.0One-click installs, shared models
stable-diffusion.cpp logo

stable-diffusion.cpp

C/C++ runtime

A ggml-based C/C++ inference engine, built like llama.cpp, with prebuilt CUDA, ROCm, Vulkan, Metal and CPU binaries, GGUF quantisation and an embedded web UI.

MITNo Python · GGUF · Vulkan
Diffusers logo

Diffusers

Python library

Hugging Face's Python library of diffusion pipelines, the code layer that SD.Next and InvokeAI build on, for when generation belongs in your own script or service.

Apache-2.0Pipelines for your own code
Mochi Diffusion logo

Mochi Diffusion

Native Mac app

A native SwiftUI app for Apple Silicon that runs Core ML Stable Diffusion models and FLUX.2 Klein on Metal, with no Python install.

GPL-3.0Core ML + Metal, macOS only

Licence, models and GPU support side by side

Code licence first. Model support lists representative recent families, not every checkpoint.

ToolLicenseInterfaceRecent modelsGPU support
ComfyUIGPL-3.0Node graph + APIFLUX.2, Qwen-Image 2.1, Ming-Image, Krea 2NVIDIA, AMD (Linux + Windows), Intel, Apple
SwarmUIMITGenerate form + Comfy tabFLUX.2, Qwen Image 2.1, Krea 2, HiDream O1As ComfyUI; multi-GPU
InvokeAIApache-2.0Layered canvas + workflowsFLUX.2, Z-Image, Krea 2, Qwen Image EditNVIDIA, AMD (Linux only), Intel Arc, Apple
SD.NextApache-2.0Gradio web UIQwen-Image 2.1, Anima, FLUX.2NVIDIA, AMD (ROCm, ZLUDA), Intel, DirectML, Apple
Krita AI DiffusionGPL-3.0Krita pluginFLUX.2, Z-Image, Qwen Image 2.1NVIDIA (6 GB+ rec.), AMD, Intel, Apple
Forge NeoAGPL-3.0A1111-style tabsFLUX.2 Klein, Z-Image, Qwen-Image, Krea 2NVIDIA first; AMD unsupported
Stability MatrixAGPL-3.0Launcher + Inference tabWhatever the installed UI supportsPer package; ROCm 10 on Windows
stable-diffusion.cppMITCLI + embedded web UIQwen-Image 2.1, FLUX.2, Z-Image, Krea2CUDA, ROCm, Vulkan, Metal, CPU
DiffusersApache-2.0Python APIMost open families as pipelinesAny PyTorch device
Mochi DiffusionGPL-3.0Native macOS appCore ML SD, FLUX.2 KleinApple Silicon only

Model weights carry their own licences: FLUX.1-dev, FLUX.2-dev and Qwen-Image 2.1 are non-commercial. Checked against each project's README and docs in September 2026.

Detailed reviews

What each tool does well, what it runs on, and when to pick something else.

Main generation UIs for current models

Start here

Four full applications that add new open model families within days of release and run on NVIDIA, AMD, Intel and Apple hardware. Pick by interface: node graph, generate form, layered canvas or a Gradio web UI with the widest backend list.

Comfy-Org/ComfyUI GitHub social previewGitHub preview
ComfyUI logo
#01

ComfyUI

Node workflow engineGPL-3.0 116k

Best for: Anyone who expects to try more than one model family this year and wants workflows they can save, share and call over an API.

When not to choose: The graph is the interface. If you want one prompt box and a Generate button, SwarmUI or InvokeAI will feel faster to learn, and every custom node you install is third-party Python running with your permissions.

ComfyUI is where new models land first. Version 0.38.0 (29 September 2026) added Ming-Image and a tiny VAE for Qwen-Image 2.1, and the core team ports memory optimisations across the FLUX and Lumina families in the same release.

Every generation is a graph of nodes, so a workflow saved as JSON or embedded in a PNG reproduces exactly on another machine. The README claims the largest open models run on 4 GB of VRAM with 8 GB of RAM through weight streaming; expect it to be slow at that size.

The official docs list CUDA on NVIDIA, ROCm 7.2 on Linux and ROCm 10.0 on Windows for RDNA 2–4, Arc via torch.xpu, and MPS on Apple Silicon. There is no official Docker image. Optional paid Partner Nodes call closed cloud models; start with --disable-api-nodes to keep everything offline.

Key strengths

  • Day-one or early support for most new open image and video models
  • Workflows saved as JSON or embedded in output images
  • Local HTTP and WebSocket API for automation
  • Runs fully offline; paid API nodes can be disabled
Sources for these claims
Desktop app, Windows portable or git + venv · NVIDIA, AMD, Intel, Apple
mcmonkeyprojects/SwarmUI GitHub social previewGitHub preview
SwarmUI logo
#02

SwarmUI

Generate-tab UIMIT

Best for: Someone who wants FLUX.2, Qwen-Image or Krea 2 working this week but would rather fill in a form than wire nodes.

When not to choose: The last tagged release is 0.9.8-Beta from February 2026; fixes ship on the master branch, so you update by pulling rather than by version. It is a .NET server plus a ComfyUI process, which is one more moving part than ComfyUI alone.

SwarmUI's Generate tab exposes the usual controls, and it builds the matching ComfyUI workflow for each model automatically. The Comfy Workflow tab gives advanced users the raw graph when the form runs out.

Its model support page is the most useful table on the topic. It lists every supported family with parameter count and a frank quality verdict, from SDXL to Qwen Image 2.1, Krea 2, HiDream O1 and Ming Image, and it notes that an fp8 model needs roughly 1 GB per billion parameters.

Beyond single images, it has a Grid Generator for parameter sweeps, video models such as Wan and LTX-2, and documented setups for multiple GPUs, Docker and AMD Radeon cards.

Key strengths

  • Automatic ComfyUI workflow for each supported model
  • Grid Generator for comparing prompts, seeds and settings
  • Multi-GPU and multi-machine backends
  • Model support table with size and quality notes
Sources for these claims
Windows, Linux or macOS installer · Docker · ComfyUI backend
invoke-ai/InvokeAI GitHub social previewGitHub preview
InvokeAI logo
#03

InvokeAI

Canvas studioApache-2.0 27k

Best for: Composing and retouching images on a canvas with SDXL, FLUX or Z-Image, where you want layers, masks and boards instead of a node graph.

When not to choose: Invoke's docs support AMD GPUs on Linux only, and new architectures tend to arrive after ComfyUI and SD.Next. Pick ComfyUI if you need a model the week it drops.

Invoke puts generation on a unified canvas with layers, brush tools, inpainting and outpainting, and keeps outputs in boards with their metadata. A workflow editor is there when the canvas is not enough.

Version 6.14.2 (27 September 2026) added custom fonts for the canvas text tool and style transfer for Krea 2. The README's model list runs from SD 1.5 and SDXL through SD 3.5, FLUX.1 and FLUX.2, Z-Image, Krea 2, Anima and Qwen Image Edit.

Its system requirements page is unusually concrete: 8 GB VRAM for SDXL, 10 GB for FLUX.1 and 12 GB for FLUX.2 Klein 4B (8 GB in FP8), with 16 GB of unified memory recommended on Apple Silicon.

Key strengths

  • Unified canvas with layers, brushes, inpainting and outpainting
  • Node workflow editor alongside the canvas
  • Per-model VRAM and RAM guidance in the docs
  • Apache-2.0 code licence
Sources for these claims
Invoke Launcher or pip/uv · localhost web UI · NVIDIA, AMD (Linux), Intel Arc, Apple
vladmandic/sdnext GitHub social previewGitHub preview
SD.Next logo
#04

SD.Next

All-in-one web UIApache-2.0 7.1k

Best for: An AMD, Intel Arc or DirectML machine where the NVIDIA-first tools keep failing, or anyone who wants reference models downloaded on first use.

When not to choose: Releases are dated updates on a rolling branch rather than semantic versions, and the interface packs in a lot of settings. It is not the place for ComfyUI custom nodes.

SD.Next lists CUDA, ROCm on Linux and Windows, ZLUDA, Intel IPEX, OpenVINO, DirectML, Apple MPS and ONNX/Olive, and tunes itself for the detected platform on install. Its 21 September 2026 update moved Windows ROCm installs to stable and added Qwen-Image 2.1.

The project claims its SDNQ quantisation cuts VRAM use by up to 4× with little quality loss, and Balanced Offload moves weights between GPU and system RAM. Selecting a reference model downloads it for you.

Key strengths

  • Eight documented compute backends, including ZLUDA and OpenVINO
  • SDNQ quantisation and balanced CPU/GPU offload
  • Automatic download of reference models
  • Built-in captioning, tagging and upscaling tools
Sources for these claims
git clone + launcher script · Docker recipes · CUDA, ROCm, ZLUDA, IPEX, OpenVINO, DirectML, MPS

Painting, A1111-style tabs and launchers

These tools are built around one way of working: generating inside Krita, keeping the AUTOMATIC1111 layout on a maintained codebase, or installing several UIs side by side with shared model folders.

Acly/krita-ai-diffusion GitHub social previewGitHub preview
Krita AI Diffusion logo
#05

Krita AI Diffusion

Painting pluginGPL-3.0 10k

Best for: An illustrator who already paints in Krita and wants generative fill, live painting and upscaling without leaving the canvas.

When not to choose: It only makes sense if you use Krita. Model choice follows the plugin's presets and the ComfyUI server it manages, so exotic workflows still need ComfyUI directly.

The plugin works with Krita's own selections and layers: select an area to fill, expand or remove objects, or turn on Live Painting to have the model reinterpret your strokes as you draw. Regions let you attach a different prompt to each layer.

It installs and manages a local ComfyUI server for you, or connects to one you already run. The README recommends an NVIDIA card with at least 6 GB of VRAM and lists ROCm, Intel XPU and MPS on macOS 14+. Qwen Image 2.1 support, including transparency, landed on the main branch on 27 September 2026.

Key strengths

  • Selection-based inpainting and outpainting
  • Live Painting with real-time feedback
  • Per-layer regional prompts
  • Upscaling to 4k and 8k without running out of memory
Sources for these claims
Krita plugin + managed or existing local ComfyUI · NVIDIA, AMD, Intel, Apple
Haoming02/sd-webui-forge-classic GitHub social previewGitHub preview
Forge Neo logo
#06

Forge Neo

A1111-style web UIAGPL-3.0 1.3k

Best for: An AUTOMATIC1111 or Forge user who wants the same tabs and extensions workflow on a project that still ships releases.

When not to choose: It is NVIDIA-first: the README says AMD issues are ignored, and the Docker setup is NVIDIA only. FLUX.2 Dev is not supported, only Klein 4B and 9B.

lllyasviel's original Forge has had no commits since June 2025. Forge Neo, on the neo branch of Haoming02's fork, picks up from Forge's Gradio 4 line and adds current models: Krea 2 Turbo and Raw, Anima, FLUX.2 Klein, ERNIE-Image, Z-Image, Qwen-Image and Qwen-Image-Edit, Chroma and Wan 2.2 14B.

Release 2.29.1 shipped on 21 September 2026. The wiki's Inference References page lists recommended settings per model, which saves a lot of trial and error with the newer architectures.

Key strengths

  • AUTOMATIC1111-style tabs; most base A1111 features still work
  • Mixed-precision formats including fp8, nvfp4 and mxfp8
  • SageAttention and FlashAttention options
  • Per-model inference settings in the wiki
Sources for these claims
git + Python 3.13 venv (uv optional) · Docker for NVIDIA
LykosAI/StabilityMatrix GitHub social previewGitHub preview
Stability Matrix logo
#07

Stability Matrix

Package managerAGPL-3.0

Best for: Trying two or three UIs on one machine without separate Python installs and duplicate 20 GB checkpoints.

When not to choose: It adds a management layer between you and each UI, and its own Inference tab is a curated front end on ComfyUI. If you only ever use one UI, install that UI directly.

Stability Matrix ships its own Git and Python, so installing a package does not touch your system environment, and the whole data folder can move to another drive. Packages share one model directory through links.

Version 2.16.4 (16 September 2026) moved Windows ROCm installs to AMD's ROCm 10 package repositories and stopped the installer from silently deleting an existing ComfyUI folder it was asked to import. The built-in Inference UI builds a ComfyUI graph for each tab and sends it to a connected ComfyUI backend.

Key strengths

  • One-click install and update for 20+ UI and training packages
  • Shared model folders across packages
  • No global Python or Git required
  • Model browser for Civitai and Hugging Face
Sources for these claims
Portable app for Windows, Linux and Apple Silicon macOS · bundled Git and Python

Lightweight runtimes, code and native Mac apps

Pick one of these when the Python web UIs are the wrong shape: a C/C++ binary for low VRAM or Vulkan, a library for your own scripts, or a signed macOS app.

leejet/stable-diffusion.cpp GitHub social previewGitHub preview
stable-diffusion.cpp logo
#08

stable-diffusion.cpp

C/C++ runtimeMIT 6.2k

Best for: A low-VRAM card, an AMD GPU on Vulkan, or any machine where the PyTorch stack will not install cleanly.

When not to choose: The README warns that the API and command-line options change often. ControlNet covers SD 1.5 only, and there is no node or extension ecosystem.

stable-diffusion.cpp runs diffusion models without Python. Releases ship Windows and Linux binaries for CUDA, ROCm, Vulkan and plain CPU, plus a macOS arm64 build, and the source also supports OpenCL and SYCL.

Model support is broad for a runtime this small: SD 1.x to SD 3.5, FLUX.1 and FLUX.2, Qwen Image, Z-Image, Krea2, HiDream-O1 and Ideogram4, with day-zero Qwen-Image-2.1 support on 20 September 2026. Weights can be quantised to q8_0, q5 or q4 on load or converted to GGUF ahead of time; the project's table puts SD 1.x at 512×512 near 2 GB.

Since April 2026 the server ships an embedded web UI, so it is usable without the CLI.

Key strengths

  • Prebuilt CUDA, ROCm, Vulkan and CPU binaries
  • GGUF and q4–q8 quantisation on load
  • Embedded web UI in sd-server
  • Video models: Wan, LTX-2 and MiniMax-H3
Sources for these claims
Prebuilt binaries (CUDA, ROCm, Vulkan, CPU, macOS arm64) · sd-server web UI
huggingface/diffusers GitHub social previewGitHub preview
Diffusers logo
#09

Diffusers

Python libraryApache-2.0 34k

Best for: A developer batch-generating product shots or wiring image generation into a Python backend, with no UI in the way.

When not to choose: There is no interface. You write the pipeline, manage memory and handle model downloads yourself, and community formats such as single-file checkpoints and LoRAs need more care than in the UIs.

Diffusers gives each model family a pipeline class you load with from_pretrained and call with a prompt. Version 0.40.0 (20 August 2026) added LTX-2.5, MiniMax-H3 and Wan Animate 2 pipelines and declared Modular Diffusers stable.

The memory guide covers CPU offload, VAE tiling and slicing, and the quantisation docs cover bitsandbytes, GGUF, torchao, Quanto and SDNQ backends. That is the same toolbox the UIs use, exposed directly.

Key strengths

  • Pipelines for most open image and video models
  • Quantisation via bitsandbytes, GGUF, torchao and Quanto
  • CPU offload, VAE tiling and slicing for low VRAM
  • Loads weights straight from the Hugging Face Hub
Sources for these claims
pip install diffusers · PyTorch on CUDA, ROCm, MPS or CPU
MochiDiffusion/MochiDiffusion GitHub social previewGitHub preview
Mochi Diffusion logo
#10

Mochi Diffusion

Native Mac appGPL-3.0

Best for: A Mac user who wants a double-click app for offline generation and does not need the newest model families.

When not to choose: macOS only, and model choice is narrow: Stable Diffusion checkpoints must be converted to Core ML, each conversion is locked to one image size, and the only modern family is FLUX.2 Klein.

Mochi Diffusion uses Apple's Core ML Stable Diffusion implementation, which can run on the Neural Engine with low memory use, and the Iris pipeline for FLUX.2 Klein 4B and 9B on Metal. FLUX.2 Klein needs no conversion.

Version 6.2 shipped on 30 September 2026. Images save as PNG with generation settings in AUTOMATIC1111-compatible metadata that Civitai uploads read.

Key strengths

  • Native macOS app, fully offline
  • Neural Engine or GPU compute unit choice
  • FLUX.2 Klein on Metal without conversion
  • A1111-compatible metadata in saved PNGs
Sources for these claims
macOS app (.dmg) · Apple Neural Engine or GPU

Decision guide: which image generation tool to install

Match your hardware and way of working to a starting point.

First install, current models

Most new open models reach ComfyUI first, and its desktop app removes the Python setup.

Start with ComfyUI; switch to SwarmUI if the graph gets in the way. Jump to review →
ComfyUI models without nodes

SwarmUI builds the ComfyUI workflow for you and keeps the raw graph one tab away.

Inpainting and composition

Layers, masks and brushes matter more than raw model count for retouching work.

Use InvokeAI, or Krita AI Diffusion if you already paint in Krita. Jump to review →
AMD or Intel GPU on Windows

SD.Next documents ROCm, ZLUDA, IPEX, OpenVINO and DirectML; Invoke supports AMD on Linux only.

Use SD.Next, or stable-diffusion.cpp on Vulkan if PyTorch will not cooperate. Jump to review →
Coming from AUTOMATIC1111

A1111 master has been frozen since July 2024; Forge Neo keeps the tabs and adds 2026 models.

Move to Forge Neo on an NVIDIA card. Jump to review →
4–8 GB of VRAM

GGUF q4–q8 weights and a runtime without PyTorch overhead leave more memory for the model.

Try stable-diffusion.cpp with a quantised model. Jump to review →

Questions people ask

Maintenance, VRAM, model licences and hardware questions that come up in forums and issue trackers.

Is AUTOMATIC1111 still maintained?

Barely. Its master branch has had no commits since 27 July 2024 and the latest release is v1.10.1; the dev branch got installer fixes in February and March 2026 but no new model support. lllyasviel's Forge stopped in mid-2025 and Fooocus is in bug-fix-only LTS for SDXL. Forge Neo is the maintained successor with the same layout.

How much VRAM do I need for SDXL, FLUX or Qwen-Image?

Invoke's requirements page lists 8 GB for SDXL, 10 GB for FLUX.1 and 12 GB for FLUX.2 Klein 4B (8 GB in FP8). SwarmUI's rule of thumb is about 1 GB per billion parameters in fp8, so a 20B model like the original Qwen-Image needs quantisation or offload on most consumer cards. GGUF and SDNQ builds bring these down further at some cost in speed or quality.

Can I sell images made with FLUX or Qwen-Image?

Check the model licence, not the UI licence. FLUX.1-dev, FLUX.2-dev and FLUX.2 Klein 9B use Black Forest Labs non-commercial licences, while FLUX.1-schnell and FLUX.2 Klein 4B are Apache-2.0. Qwen-Image is Apache-2.0, but Qwen-Image 2.1 ships under the Qwen Research License for non-commercial use, and SD 3.5 is free commercially only below USD 1M annual revenue.

Does local image generation work on AMD GPUs?

Yes, with more friction than NVIDIA. ComfyUI documents ROCm 7.2 on Linux and ROCm 10.0 on Windows for RDNA 2–4, SD.Next adds ZLUDA and DirectML, and stable-diffusion.cpp ships Vulkan and ROCm binaries. InvokeAI supports AMD on Linux only, and Forge Neo does not handle AMD issues.

What is the best option on a Mac?

ComfyUI, InvokeAI and SD.Next run on Apple Silicon through PyTorch MPS; ComfyUI does not use MLX. Mochi Diffusion is the open-source native app. Draw Things is popular, but its App Store app is closed source; only its community engine repository is GPL-3.0.

Are ComfyUI custom nodes safe to install?

Treat them as untrusted code. A custom node is Python that runs with your user's permissions and can install dependencies, so read the source or stick to well-known repositories, and keep the server off the public internet.

Primary sources