On Windows with a compatible NVIDIA setup, the shortest documented route is the official one-click package: extract it, run update.bat, then run.bat. Choose the Git route if you need control over the environment or are using newer hardware.
Download only from lllyasviel/stable-diffusion-webui-forge. We do not mirror installers. The release files and source links on this page lead to the original GitHub repository.
Best for a clean Windows install when your NVIDIA GPU and driver work with the package’s CUDA 12.1 / PyTorch 2.3.1 environment. Git and Python are included.
Recommended in the original READMEOpen the complete Windows guideUse this route when you want visible source history, control over Python and packages, or need to solve compatibility for hardware newer than the published archive.
Requires Git, Python, and troubleshooting comfortOpen the complete Git guideVerify the NVIDIA driver, Git, Python 3.10 and venv, then clone the original repository and let webui.sh create its local environment. The README does not publish this as a supported parity route.
Native Linux and WSL 2 have different driver boundariesOpen the complete Linux guideOriginal Forge has AMD-related source paths but no current AMD support matrix or one-click package. Check the exact GPU and operating system against ROCm or test DirectML only in a disposable clone.
Source-audited routes · no universal AMD support claimOpen the complete AMD guideOriginal Forge contains macOS defaults and MPS code but no current Mac support matrix. Check Apple Silicon, macOS, native Python and a standalone MPS tensor before using a disposable Forge clone.
Source-audited lab · no Mac hardware generation claimOpen the complete macOS guidePackage managers and community procedures have their own owners, update logic and support boundaries. Do not relabel them as original Forge installers.
No parity claim with Windows + NVIDIAReview the platform boundaryThe package removes the need to install Git and Python separately. It does not make every GPU, driver, operating system, or model compatible.
Users have reported that the archive’s bundled PyTorch does not recognize newer sm_120 GPUs such as the RTX 50 series. That is a community report, not an official compatibility matrix. If the console says your GPU architecture is unsupported, use a controlled environment with a compatible current PyTorch build; do not hide the test with a skip flag.
Inspect the reported RTX 50-series issue ↗The project calls this a one-click package because Git and Python are included. You still extract the files, prepare the environment, and keep the local console running.
Read the original README ↗Verify the owner is lllyasviel and the repository name is stable-diffusion-webui-forge. The release tag is named latest; its three Windows assets were uploaded on 11 August 2024.
The original README recommends the CUDA 12.1 / PyTorch 2.3.1 archive. “Newer” is not automatically safer for this release.
| File | Environment | Project note | Use |
|---|---|---|---|
webui_forge_cu121_torch231.7z | CUDA 12.1 · PyTorch 2.3.1 | README recommendation | Start here for the documented Windows package path. |
webui_forge_cu121_torch21.7z | CUDA 12.1 · PyTorch 2.1 | Older environment | Kept as the previously used package; not the default choice. |
webui_forge_cu124_torch24.7z | CUDA 12.4 · PyTorch 2.4 | Compatibility warning | README calls it fastest but warns MSVC may be broken and xformers may not work. |
The release API did not publish checksums for these assets when checked. We do not invent or re-host hashes.
Use 7-Zip or another 7z-compatible extractor. Open the extracted folder—not the file preview inside the archive. You should see the system and webui folders plus the batch files.
update.bat firstThis is an explicit step in the README. Let the console finish and read any error before closing it. If Windows asks whether to run the batch file, verify that you extracted the original GitHub asset.
update.batrun.bat and keep the console openForge prepares the environment, starts a local server, and normally opens the browser UI. Success means the console remains running and prints a local address. Closing that console stops the interface.
run.batThis route exposes the repository and environment directly. It is better suited to experienced users, newer GPUs, and installations that need deliberate package versions.
The inspected launcher warns when Python is outside its tested range and names Python 3.10.6 as the tested version. Treat that as a snapshot of the original code, not a promise that every package combination will remain available.
Inspect the launcher code ↗Confirm both commands resolve in the terminal before cloning. Keep this Python environment separate from unrelated AI tools.
git clone https://github.com/lllyasviel/stable-diffusion-webui-forge.gitcd stable-diffusion-webui-forgeOn Windows, run webui-user.bat. The file leaves Python, Git, virtual-environment, and command-line overrides empty by default so advanced users can set them explicitly.
webui-user.batIf you already use AUTOMATIC1111, Forge can be configured to reference existing model directories. Do not point two installations at the same virtual environment; reuse model paths only after both installations work independently. Follow the reversible A1111 → Forge migration protocol.
A quiet browser does not mean setup has stalled. The console is the source of truth while the environment is being built.
Forge prepares its Python environment.
Torch and required packages are checked or installed.
CLIP, OpenCLIP, and repository assets may be downloaded.
The running console prints the local browser address.
Keep the console open, copy the first traceback, and record the route, GPU, driver, Python version, and package filename before changing anything.
Use a separate compatible 64-bit environment. A system Python upgrade can break an existing virtual environment even when Forge files did not change.
Verify the NVIDIA driver and the PyTorch build together. An RTX 50-series architecture error is not fixed by suppressing the CUDA test.
Check current PyTorch options ↗Preserve the exact URL and error. Retry only after checking connectivity, proxy, certificate, antivirus, and free disk space; repeated partial installs can obscure the cause.
See a community dependency report ↗Do not scatter global pip install commands across environments. Recreate or repair the Forge environment with compatible pinned packages.
Return to the console. If the process exited, localhost has nothing to connect to. If it is running, use the exact local address it printed and check whether security software blocked the process.
The application is installed; the model is missing or in the wrong folder. Continue with the first-run guide instead of reinstalling Forge.
Set up the first model →Linux, AMD, macOS, and package-manager instructions can be useful. They must still be validated against the current distribution, Python packages, driver, and acceleration backend.
The repository contains a Unix launcher, while installation guidance is spread across code, an open issue and an unmerged documentation PR. Our NVIDIA guide separates native Linux from WSL 2 and tests every prerequisite before cloning.
ROCm and DirectML routes depend on operating system, exact GPU, backend and PyTorch versions. The original README does not provide a current AMD support matrix or equivalent one-click archive. Our AMD guide separates native Linux ROCm, ROCm in WSL 2 and a reversible Windows DirectML experiment.
Choose an AMD compute pathOriginal Forge contains a macOS launcher and MPS branches, but it does not publish a current Mac support matrix. The dedicated guide separates Apple Silicon/MPS readiness from Forge compatibility and starts with a reversible clean-clone test.
Tools such as Stability Matrix can manage UIs and shared models. They are separate products with their own update logic and support boundaries, not the original Forge installer.
Open Stability Matrix ↗These are the checks we use to separate an installation problem from a first-run, model, or compatibility problem.
For the documented Windows + NVIDIA route, the original README recommends webui_forge_cu121_torch231.7z. Check your GPU compatibility first: the package was uploaded in August 2024 and its PyTorch build may predate newer GPU architectures.
Not for the one-click Windows package; the README says both are included. You do need them for the Git route, where version compatibility becomes your responsibility.
The packaged PyTorch environment brings the CUDA runtime it was built to use, so most users do not need a separate developer toolkit. You still need a compatible NVIDIA driver. If you build packages from source, toolkit requirements can change. Check NVIDIA’s compatibility documentation for the driver/runtime relationship.
No model is promised in the original package. Install Forge first, confirm the UI starts, then add one compatible checkpoint and its required companion files.
Yes. Keep the repositories and virtual environments separate. After both work, advanced users can point Forge to existing model directories instead of duplicating large files. Use the complete A1111 migration guide to inventory settings, extensions and the return path.
Back up custom configuration first. Package users follow the package’s update.bat workflow; Git users inspect changes and update the repository deliberately. If the environment is already broken, capture the error before updating so the cause is not erased.
The GitHub release metadata did not expose asset digests when we checked it. Download from the original repository, verify the owner and release URL, and do not trust hashes copied from an unrelated mirror.
Do not assume it will. Community reports show sm_120 architecture errors with the package’s older PyTorch environment. Check current PyTorch support for your exact GPU and use the controlled Git route when the bundled environment is too old.