How to Install Forge Neo on Windows: The Steps for Stable Diffusion WebUI Forge and Where the First Launch Gets Stuck

What Forge Neo is and where to get it: start with the repo name and the branch

In Part 2, Alex installed ComfyUI Desktop. Wiring nodes together made it easy to save a workflow. But for typing one prompt and pulling several images while changing only the size and step count, a screen full of input fields and buttons felt more natural. So Alex decided to add Forge Neo, which uses the A1111-style interface.

The name is the first thing that trips people up. The official Forge Neo repository lives at github.com/Haoming02/sd-webui-forge-classic. The repository name says "classic," but its default branch is neo, and the README on that branch is titled "Stable Diffusion WebUI Forge - Neo." The same repository also has a separate classic branch, and the top of the README links to Neo and Classic separately.

The README describes Neo as a continuation of the "latest" version of Forge, the project lllyasviel built on top of the AUTOMATIC1111 WebUI. Its stated focus is optimization and usability, with the goal of running recent popular models through an easy-to-use interface. The feature list opens by saying that most base features of the original A1111 WebUI should still work.

Last activity in the three repositories

You can confirm that A1111 has stalled by looking at the repository history. On November 6, 2026, the GitHub records of all three projects showed the following. The last commit on A1111's default branch (master) is a changelog commit from July 27, 2024, for v1.10.1. The last merge on its development branch (dev) is an install fix from March 2, 2026, and there has been no new release after v1.10.1. The last commit on the original Forge main branch is from June 26, 2025. Forge Neo published release 2.31 on October 21, 2026, and the neo branch received commits as recently as November 5, 2026.

The three install methods are different, too

A similar screen does not mean a similar install. Here is how the three READMEs describe a Windows install, side by side.

The row that affects the install right away is Python. Even if your PC already has Python 3.10 from an older guide, Neo will not simply reuse it. The next section explains why.

The one-click package that the original Forge README recommends does not exist for Neo. Release 2.31 has no attached files, and the README's install steps start with git clone. If you unzipped a "Forge one-click package" from the internet, what you have is the original Forge.

What to prepare before installing: git, uv and your graphics driver

The README's install steps begin with "install git." After that, Python setup splits into two paths. The one the README lays out as the "Recommended Method" uses uv. Installing Python 3.13.12 directly is tucked away as the "Deprecated Method." Alex prepared three things.

FFmpeg is optional. The README says you need FFmpeg installed to export videos. If you only plan to make images, you can skip it for now.

You do not need to install Python yourself: uv fetches 3.13

The heart of the recommended method is one line: uv venv venv --python 3.13 --seed. The uv docs explain that, by default, uv automatically downloads a Python version if it cannot find it on the system. So even if your PC has no Python, or only a different version, this command creates a Python 3.13 virtual environment named venv inside the Neo folder. You do not have to install Python 3.13 system-wide. The trailing --seed installs seed packages such as pip into the environment.

The reason the version matters is in the launch script. Neo's modules/launch_utils.py checks the Python version at startup. If the minor version is below 13, it warns that the program was tested with Python 3.13.15. The warning goes on to say that if you run into package install errors, you should upgrade to the latest 3.13 release and delete the venv folder in the WebUI directory. It does not block the launch, but it does mean a virtual environment built on 3.10 or 3.12 may fail during package installation.

Decide on the install folder and drive first

git clone creates a new folder under whatever folder you run it from. With the README command, that folder is named sd-webui-forge-neo. The official docs do not prescribe a location; this guide uses C:\AI as an example.

Leave plenty of disk space. When the README describes the --uv-symlink option, it says the option cuts the install size from about 7GB to about 100MB. That suggests a default install, which copies packages, takes around 7GB. Models come on top of that. The SDXL base checkpoint alone is a single 6.94GB file, as listed on its Hugging Face page.

If your C: drive is tight, clone to another drive and add --uv-local-cache. The README says this option keeps the uv cache in a .uv-cache folder inside the WebUI directory, which speeds up installs on drives other than the system drive and lets you uninstall cleanly by deleting the WebUI folder.

Forge Neo install steps: five steps from clone to first launch

Once the prerequisites are in place, you only type a few lines. Here are the README's install steps, written for the Windows Command Prompt.

mkdir C:\AI
cd /d C:\AI
git clone https://github.com/Haoming02/sd-webui-forge-classic sd-webui-forge-neo --branch neo

cd sd-webui-forge-neo
uv venv venv --python 3.13 --seed

Of the five steps, you type commands in the first two and edit a file in the third. webui-user.bat downloads and installs everything else.

Not on Windows? AMD, Linux and macOS

The README points AMD users to CS1o's guide (installing Forge Neo with ROCm), Linux and macOS users to the Unix page of the repository wiki, and Docker users to the repository's docker folder. The Unix wiki follows the same outline: install git, FFmpeg and uv, clone, run uv venv venv --python 3.13 --seed, then run ./webui-user.sh. On macOS, though, the wiki says to uncomment the TORCH_COMMAND line so you get regular PyTorch instead of the CUDA build, and it warns that macOS does not support any of the attention packages. The README's issue guidelines also state that AMD GPU issues will not be answered.

webui-user.bat: where options go and what to put there

Almost every setting that changes the install lives in one file, webui-user.bat. Right after cloning, it looks like this.

@echo off

:: set PYTHON=
:: set GIT=
:: set VENV_DIR=

set COMMANDLINE_ARGS=

:: --xformers --sage --uv
:: --pin-shared-memory --cuda-malloc --cuda-stream
:: --skip-python-version-check --skip-torch-cuda-test --skip-version-check --skip-prepare-environment --skip-install

call webui.bat

Lines that start with :: are comments and do not run. The comment lines near the bottom are samples of commonly used options. As the README says, put options on the same line after set COMMANDLINE_ARGS=, separated by spaces. The README adds that you can see the full list by running python launch.py --help in the Neo folder. These are the options that matter for installation.

Attention options install packages; they do not switch them on

This is the easiest part of the README to misread. --xformers, --flash and --sage only install packages. They do not decide which attention is used, and you can remove them once the install is done. At startup, Neo tries the installed packages and picks the first one that works, in this order: SageAttention, FlashAttention, xformers, PyTorch, Basic. To skip a specific package, use its disable flag, such as --disable-sage.

The README warns against installing all of them blindly and notes that PyTorch's built-in attention (scaled_dot_product_attention) is usually just as fast and more stable. It also states that xformers does not support RTX 50 series cards. For a first install, start with --uv alone and add other options after your first image comes out. That way, if something breaks, it is easier to tell what caused it.

If you already have models downloaded

If your models already live in another app, you do not need to download them again. Per the README, --forge-ref-a1111-home points to an A1111 install folder and --forge-ref-comfy-home points to a ComfyUI install folder, and Neo then reads their model folders. --model-ref works differently. It points Neo at a central models folder that replaces Neo's own models folder; in the README's words, it replaces rather than adds. How to keep one model library that several apps share is covered in sharing one model folder across several apps (in Korean).

What happens on the first launch: reading the console line by line

The first launch is the other half of the install. When you run webui-user.bat, text scrolls through the console for a while. If you follow the launch scripts (webui.bat and launch_utils.py), the order is fixed. Knowing that order is the fastest way to tell where things stopped.

1. Activate the virtual environment: it finds and activates the Python inside venv. If neither uv nor Python can be launched, it stops right here.
2. Print versions: if Python is not 3.13 or newer, the version warning appears first. Then you see lines such as Python 3.13… and Version: neo 2.31.
3. Install PyTorch: if torch is missing, it prints "Installing PyTorch" and downloads the CUDA 13.0 builds of torch 2.14.0 and torchvision 0.29.0 from the official PyTorch package index.
4. Check the GPU: it checks whether the installed PyTorch can use a CUDA device (or an Intel XPU or Apple MPS device). If not, it stops here.
5. Install the remaining packages: it installs the packages in requirements.txt, plus Gradio 4.40.0, which draws the interface. If you added --uv, these installs run through uv.
6. Open the interface: the WebUI server starts and your browser opens. The default port is 7860.

The browser opens on its own because the "Launch the webui in browser on startup" setting under Settings → System defaults to Local. If the browser does not open, copy the local address printed in the console and open it yourself.

From the second launch on, step 3 is much shorter. The launch script skips the PyTorch install if torch and torchvision are already installed, and only reinstalls them when you add --reinstall-torch.

"No model" on the first screen is normal

Neo does not ship with model files. If there is no checkpoint at all, the console prints You do not have any model... Please download models to "models/Stable-diffusion". That does not mean the install failed. It just means the model folder is empty. The interface is already running, so add a model and refresh the list.

Where models go: the models folder and UI Presets

When you clone Neo, the models folder already contains a subfolder for each model type, and each one holds a small text file with a hint such as "Put Checkpoint here."

models
├── Stable-diffusion   checkpoints / main diffusion models (.safetensors · .ckpt)
├── text_encoder       text encoders
├── VAE                VAEs
├── Lora               LoRAs
├── ESRGAN             upscalers (.pth · .safetensors)
└── embeddings         embeddings

An SD 1.5 or SDXL checkpoint packs everything into one file, so dropping it into Stable-diffusion is enough. Models like FLUX, where the main model, text encoder and VAE come as separate downloads, go in three places: the main model in Stable-diffusion, the rest in text_encoder and VAE. The Neo interface has a checkpoint box and a separate "VAE / Text Encoder" box. The second box shows files from both the VAE and text_encoder folders in one list, so you can select several at once. The "Download Models" page of the repository wiki lists which files to download for each model family.

The README specifically says to put every upscaler, whatever its format, in the ESRGAN folder. By default, generated images pile up under output\txt2img-images in the Neo folder. If you added a file and it does not show up in the list, click the refresh button next to the checkpoint box.

Pick the model family with UI Preset

The Neo interface has a "UI Preset" selector with model families such as sd, xl, flux, klein, qwen, qwen21, lumina, zit, wan, anima, ernie, pid, krea and moss. Each preset has its own default sampler and step count. In the code, sd defaults to Euler a with 32 steps, xl to Euler a with 24 steps, and flux to Euler with 20 steps. The README says the preset system saves checkpoint, module and parameter selections separately for each preset. When you pick an SDXL checkpoint, set the preset to xl as well.

Not every model works

Whether the model you chose for your VRAM in Part 1 runs in Neo is something to check in the README's feature list and the wiki.

GGUF support also appears in the README's Removed Features list, so plan on .safetensors files. How model size relates to VRAM is covered in which models each VRAM size can run (in Korean), and where to download checkpoints and which file format to choose is covered in where to get model files and how to pick a format (in Korean).

Where a first launch gets stuck: find it by the message

When the install stops, read the last few lines in the console. These are the messages you can hit on a first launch, taken from the wording in the launch scripts.

These six messages are not equally serious. Four of them stop the install on the spot. The Python version warning lets the launch continue. The last one is closer to a sign that the install has finished.

Switching to an older PyTorch

The wiki's "Older PyTorch" section tells you to add one line to webui-user.bat, above the call webui.bat line.

set TORCH_COMMAND=pip install torch==2.10.0+cu126 torchvision==0.25.0+cu126 --extra-index-url https://download.pytorch.org/whl/cu126

One more thing to watch. The launch script skips the PyTorch install step when torch is already installed. If the CUDA 13.0 build already went in during the first launch, adding this line alone changes nothing, so run once with --reinstall-torch. If you would rather delete the venv folder, start again from the uv venv command in Step 2.

Running out of memory: check GPU Weights first

If generation stops with an out-of-memory error, look at the settings before adding options. The README's optimization list includes a rewritten backend (memory_management.py and others) and unloading all models when an out-of-memory error occurs. The install steps do not include any VRAM option. The manual control is the "GPU Weights" slider under Settings → System. It sets the share of VRAM Forge can use, and its default is 1 (all of it). --lowvram still exists as an optional flag in the backend arguments, but it is not required. How models use VRAM is covered in Part 1.

If the window closes silently or ends at "Press any key"

The Neo FAQ (issue #414) links the "Press Anything to Continue..." symptom to discussion #1474 in the original Forge repository. That post blames several symptoms on too little system swap (Windows virtual memory): the model loads and then nothing happens, the console closes in a flash with no error, or model loading takes more than a minute. It notes that Windows 10 and 11 normally manage swap automatically, so instead of changing that setting, keep at least 40GB free on each disk. It also says model loading gets much slower when swap sits on an HDD. If you do change the swap settings, restart the computer for them to take effect.

Updating, rolling back and uninstalling, plus Alex's install

Neo is a git repository, not a zip package, so you update it as a repository. The update method the Neo FAQ links to is the wiki's GitHub Desktop guide.

• Add: if you already cloned the folder, choose it in GitHub Desktop with File → Add local repository.
• Update: click Fetch origin. If there are new commits, the button changes to Pull origin; click it once more. From the command line, this is the same as running git pull in the Neo folder.
• Roll back: if something breaks after an update, open History, right-click the last commit that worked, and choose Checkout commit. To get updates again, pick your original branch under Current branch.
• Branches: the wiki says to choose "Bring my changes" if you switch branches after editing webui-user.bat. As noted earlier, do not switch between classic and neo.

Extensions and install managers

Some extensions written for A1111 or the original Forge do not work in Neo. For those cases, the Neo FAQ links to a list of replacement extensions (discussion #1128, which includes sd-dynamic-prompts and ADetailer). The README says issues with third-party extensions will not be answered. The same goes for problems caused by install managers such as StabilityMatrix; the README asks you to report an issue only if it reproduces on a clean install that follows the official instructions. That is why, when an install goes sideways, a fresh install in a different folder using the README steps is the most reliable way to check.

How to uninstall

Neo is not installed through an installer, so it does not appear in the Windows "Installed apps" list. By default, the virtual environment, models and generated images all sit inside the cloned folder, so deleting that folder removes them. The uv cache is the exception. The help text for --uv-local-cache describes it as using a local cache (.uv-cache) "instead of the system-wide cache." If you installed without that option, the cache stays outside the folder. If you moved models elsewhere with an option such as --model-ref, that folder stays too.

Alex's Forge Neo install, and what comes next

Alex followed the README's recommended method. After installing git and uv, Alex found the C: drive short on space and, like the ComfyUI install in Part 2, cloned to the D: drive as D:\AI\sd-webui-forge-neo. Only two options went into webui-user.bat: --uv --uv-local-cache. The goal was a faster install on a non-system drive, plus a cache that disappears along with the folder if it is ever deleted. Attention packages could wait until after the first image.

The first launch downloaded PyTorch and the other packages, then opened the browser, and the console printed the line saying there was no model. Alex put one SDXL checkpoint into models\Stable-diffusion, refreshed the list, set UI Preset to xl and generated the first image. Making Neo read the ComfyUI model folders was left for after that first image.

FAQ

Where do I download Forge Neo? Is there a one-click installer?

The official repository is Haoming02/sd-webui-forge-classic on GitHub, and its default branch, neo, is Forge Neo. As of November 6, 2026, the latest release, 2.31, has no attached installer files, and the README's install steps begin by cloning the neo branch with git. The one-click zip package is the install method of the original Forge repository by lllyasviel.

Which Python version does Forge Neo need?

Python 3.13. The README's recommended method is to install uv and create the virtual environment with uv venv venv --python 3.13 --seed, and uv downloads 3.13 on its own if your PC does not have it. If the Python version is older than 3.13, the launch script warns that it was tested with 3.13.15 and tells you to move to the latest 3.13 and delete the venv folder if package installs fail.

After the first launch it says "You do not have any model." Did the install fail?

No. Neo ships without model files, so it prints this line when there is no checkpoint. Put a checkpoint in the models\Stable-diffusion folder and click the refresh button in the interface, and it shows up in the list. For split models such as FLUX, put the text encoder and VAE in the text_encoder and VAE folders.

I have an RTX 50 series card. Should I add --xformers?

No. The README states that xformers does not support the RTX 50 series. Also, --xformers, --sage and --flash only install packages, and Neo automatically picks the first usable one among the installed packages in the order SageAttention, FlashAttention, xformers, PyTorch. The README notes that PyTorch's built-in attention is usually just as fast and more stable.

My GPU has little VRAM. Do I need --lowvram?

The install steps do not include any VRAM option. The README's optimization list includes a rewritten memory management backend and unloading all models when memory runs out. If you run short, adjust GPU Weights (default 1) under Settings → System first. --lowvram is still available as an optional flag, but it is not required.

References

All sources below were checked November 6, 2026.

1. Haoming02. sd-webui-forge-classic: README (neo branch). Neo overview, features and removed features, Commandline, Installation, Attention Functions, xformers and RTX 50, uv options and install size, issue guidelines. github.com
2. Haoming02. sd-webui-forge-classic: Releases and Commits. Release 2.31 (2026-10-21, no attached files), latest neo branch commits (2026-11-05). Releases · Commits
3. Haoming02. webui-user.bat, webui.bat, modules/launch_utils.py, modules/cmd_args.py, backend/args.py, modules/shared_options.py, modules_forge/presets.py, modules_forge/main_entry.py, modules_forge/forge_version.py. Default option lines, virtual environment checks, Python 3.13 check message, torch 2.14.0 CUDA 13.0 install command, GPU check error messages, option help text, GPU Weights and browser auto-launch defaults, preset defaults, the "no model" message, version 2.31. webui-user.bat · launch_utils.py
4. Forge Neo Wiki. Extra Installations, Download Models, Unix. Installing git and uv, GitHub Desktop updates and rollbacks, the classic/neo branch warning, Older PyTorch, files per model family, Linux and macOS installs. Extra Installations · Download Models · Unix
5. Haoming02. FAQ (issue #414). How to update, replacement extension list, "Press Anything to Continue." github.com
6. lllyasviel. stable-diffusion-webui-forge: README and discussion #1474. One-click package (CUDA 12.1 + PyTorch 2.3.1), main branch commit history, System Swap (40GB free space, restart). README · #1474
7. AUTOMATIC1111. stable-diffusion-webui: README and Commits. Windows install (Python 3.10.6), last master commit (2024-07-27, v1.10.1), dev branch merge (2026-03-02). github.com
8. Astral. uv: Installation, Python versions, CLI reference. Windows install script and winget, automatic download of missing Python versions, uv venv --seed. Installation · Python versions · CLI
9. Stability AI. stable-diffusion-xl-base-1.0. sd_xl_base_1.0.safetensors, 6.94GB. huggingface.co