A local AI image generator for Windows. Everything runs on your own machine — no account, no subscription, no cloud, nothing sent anywhere unless you ask for it.
Installer (47 MB) or portable zip (57 MB) · Windows 10/11, 64-bit · free and MIT licensed
ImageForge-0.1.29-demo.mp4
21 seconds, no narration: a prompt typed live, the render on a local GPU, the gallery, and the API an agent calls. Every picture in it was made by ImageForge on an RTX 3060.
Read the guide — what every screen does, which model to use for what, training your own subject, and connecting an agent.
| Write a prompt, get a picture | Thirteen models built in, from a 3-second draft model to full-quality SDXL and FLUX. It picks sensible settings; every control explains itself in plain words rather than jargon. |
| Edit what you made | Img2img, inpainting, and structural control — draw a box, give it a pose, keep a composition. |
| Train it on your own subject | Point it at a folder of photos and it trains a LoRA adapter, captioning the images for you. |
| Runs offline | Once the engine and a model are downloaded, unplug the network and it keeps working. |
| Also an API and an MCP server | The same engine behind an HTTP API and a Model Context Protocol server, so Claude or another agent can generate images through it. |
- Download the installer —
ImageForge-*-setup.exe - Run it. No administrator needed; it installs for your user only.
- Open ImageForge. The System screen offers the image engine.
Nothing to sign up for, and nothing to configure. No account, no API key,
no administrator prompt, no firewall rule — the app's own API listens on
127.0.0.1 only, which Windows does not ask about. A HuggingFace token and a
RunPod key are optional, and only if you want gated models or cloud
rendering; the System screen's Run the checks button tells you what is
present and what each missing thing would cost you.
The app is 47 MB. The image engine is about 5 GB more, and it is not bundled — the right build depends on your graphics card, which an installer cannot know in advance.
On this machine it took 5 minutes. If you have an NVIDIA card it installs the accelerated build and says so; if you don't, it installs the processor build and warns you that renders will be much slower. It does not refuse either way.
Prefer not to install anything? The zip is the same application — unzip it
anywhere and run ImageForge.exe.
It is not code-signed, and you will see this:
Windows protected your PC Microsoft Defender SmartScreen prevented an unrecognised app from starting.
Click More info, then Run anyway.
Why, honestly: a code signing certificate costs money every month, and — this is the part that surprised me — it would not remove the warning either. Microsoft's own documentation says SmartScreen reputation "builds up automatically. The prompt stops appearing once the file hash has sufficient download history." A signed app from a publisher nobody has heard of shows the same dialog until enough people have downloaded it. Paying would buy a slower version of the same warning.
So: it's unsigned, that's a deliberate choice, and here is how to check the download is really what I published rather than taking my word for it.
Every release includes latest.json with the SHA-256 of the archive:
Get-FileHash .\ImageForge-0.1.7-win64.zip -Algorithm SHA256Compare it with the sha256 field in
latest.json. If they match, the
file is byte-for-byte what was built.
ImageForge is also an MCP server, so Claude, Cursor or any MCP client can generate images through it. The Agents screen prints the configuration to paste in — it names the engine's own interpreter and a working directory, so it works as-is once the engine is installed.
Seven tools are exposed: generate_image, edit_image, inpaint_image,
assist_prompt, list_models, rate_output and generation_insights. Each
model in list_models carries its licence, so an agent asked for something
commercial can see that SDXL-Turbo is not licensed for it.
There is an HTTP API on 127.0.0.1:8765 as well, with the same engine behind
it. Both are local only.
The app is MIT — the models are not. ImageForge ships no model weights; it downloads them from their publishers to your machine, and their licences govern what you may do with what you generate. They differ, and not in the way you would guess:
| Model | Commercial use |
|---|---|
| SD-Turbo (the default) | Yes — below US$1,000,000 of annual revenue |
| SDXL-Turbo | No. Research and personal use only |
| SDXL family, RealVisXL | Yes, with use restrictions |
| FLUX.1 schnell, FLUX.2 klein | Yes, unrestricted (Apache-2.0) |
The System screen lists every one of these with a link to the licence text, and says which of them your current model falls under. It also tells you when a feature pulls extra weights on terms of their own — inpainting and depth control both do.
See TERMS.txt for acceptable use, and
THIRD-PARTY-NOTICES.txt — shipped with the app — for
the 50 open-source packages inside it.
Nothing leaves your machine unless you ask:
- No telemetry, no analytics, no crash reporting, nothing sent at startup
- Downloading the engine or a model contacts HuggingFace, because that is where the files are
- The update check contacts GitHub only when you press it
- Your prompts and images stay on your disk
Renders go to %LOCALAPPDATA%\ImageForge\outputs. Uninstalling leaves them, and
leaves the engine, so reinstalling doesn't mean downloading 5 GB again.
| OS | Windows 10 or 11, 64-bit |
| Graphics | An NVIDIA card with 8 GB or more is strongly recommended. Works without one, much more slowly — the app installs a processor build and says so. |
| Disk | ~5 GB for the engine, plus 2–7 GB per model |
| Memory | 16 GB recommended |
Open an issue — bug reports and questions both welcome.
This repository holds downloads only. It contains no source code.