Stable Diffusion usually stops working because the launch environment is wrong: webui-user.bat closes instantly, Torch cannot see your GPU, a python module is missing, or a model will not load. Read the console message, fix the matching cause (python version, CUDA enabled Torch, a clean venv, or the right model folder), and it starts again.
Stable Diffusion is the umbrella: the underlying models plus the front ends people run them in, most commonly Automatic1111 WebUI. When it “will not work” the fault is almost always in the launch environment (python, Torch, dependencies, or model files) rather than in your prompt. The good news is that the console window prints the exact reason nearly every time. This guide is organized by what the console tells you, from the batch file that closes in a blink to the model that refuses to load.
If you run a different front end, the fix is often different too. For the node based app, see ComfyUI not working; for the Forge fork, see Forge WebUI not working. Both share DNA with Automatic1111 but have their own quirks, and this article focuses on the general Stable Diffusion and Automatic1111 layer.
Read the console first
The black command window that opens when you launch is the single most valuable diagnostic you have. Do not close it, and do not troubleshoot from the browser alone. When something fails, scroll to the last several lines and look for “Error”, “Traceback”, “RuntimeError”, or “No module named”. Every fix below is chosen based on that line. If the window closes too fast to read, launch it from an already open terminal so the text stays on screen after the crash, or add a pause so it waits before closing.
| Console message | Meaning | Section to jump to |
|---|---|---|
| Window flashes and closes instantly | Launch crashed before the UI loaded | webui-user.bat closes instantly |
| “Torch is not able to use GPU” | Torch cannot reach CUDA | Torch cannot use GPU |
| “Torch not compiled with CUDA enabled” | A CPU only Torch got installed | Torch cannot use GPU |
| “No module named …” | A python dependency is missing | No module named errors |
| “connection errored out” or endless connecting | Server not fully up or a proxy issue | Connection errored out |
| Model dropdown empty or safetensors load fails | Model in wrong folder or corrupt | Model will not load |
| Grey or blank UI in the browser | Assets blocked or server crashed | Blank or grey UI |
webui-user.bat closes instantly
A batch file that opens and vanishes means the launch crashed before the UI came up, and the error scrolled past too fast to read. The first job is to see the message. Launch webui-user.bat from an already open command prompt, or temporarily add a pause line at the end of the batch file, so the window stays open and shows the actual error. Once you can read it, it will point to one of the causes below.
The most common reason for an instant close is a python version mismatch. Automatic1111 expects a specific python (the 3.10 series is the safe, well supported choice in 2026); newer python releases frequently break dependency installation and cause an immediate exit. If you have multiple python versions installed, the launcher may be grabbing the wrong one. Point it at a compatible python explicitly, or install the recommended version and set it in the batch file. A second common cause is a half finished first run: the initial launch downloads and installs a large environment, and if that was interrupted the venv is broken. Delete the venv folder and relaunch so it rebuilds cleanly. A third is antivirus quarantining a file mid install, which leaves the environment incomplete; whitelist the folder and let it reinstall.
Torch cannot use GPU
Two closely related messages appear here: “Torch is not able to use GPU” and “Torch not compiled with CUDA enabled”. Both mean the same practical thing, that the Torch build in your environment cannot talk to your GPU, so generation would fall back to painfully slow CPU or fail outright.
The usual root cause is that a CPU only build of Torch got installed, either because the first run picked the wrong wheel or because a manual pip install pulled the default CPU package. The fix is to reinstall Torch with the CUDA enabled build that matches your setup, into the same venv the WebUI uses. Delete the existing Torch from that environment first so the CPU build does not linger. If you are on an NVIDIA card, make sure your GPU driver is current, since a very old driver can also produce this error even with the right Torch. A quick way to force a clean rebuild is to delete the venv folder entirely and relaunch, which reinstalls Torch fresh; just confirm the install command targets the CUDA build.
AMD cards are a special case, because standard CUDA Torch will never see an AMD GPU. If you run Radeon hardware you need a different path entirely, covered in the dedicated guide for Stable Diffusion on AMD GPUs. Do not keep reinstalling CUDA Torch on an AMD card; it cannot work.
No module named errors
A “No module named” traceback means a python dependency is missing from the environment. This almost always comes down to the venv: either it did not finish installing, or you installed something into your system python instead of the WebUI’s own environment. The clean fix is to let the WebUI rebuild its environment. Delete the venv folder and relaunch; the launcher reinstalls everything it needs. If you prefer to fix in place, activate the WebUI’s venv and reinstall requirements rather than installing globally, because packages in your system python are invisible to the isolated environment the WebUI actually runs in.
If a single named module keeps failing to install, it may need a build tool your system lacks, or its version clashes with your python. Confirm you are on the supported python series first, since a mismatched python is the most common reason a specific package will not install. Reinstalling into the correct, supported environment resolves the large majority of these.
xformers issues
xformers speeds up attention and saves memory, but it is version sensitive and a frequent source of startup errors, because an xformers build must match your Torch and python versions. If the console throws an xformers error at launch, the quickest diagnostic is to launch without it by removing the xformers flag, which tells you immediately whether xformers is the culprit. If the WebUI then starts fine, reinstall a version of xformers that matches your current Torch, or simply run without it, since modern Torch attention is fast enough that xformers is optional rather than essential. Never mix an xformers build meant for one Torch version with a different Torch; that mismatch is the usual cause of the crash.
Model will not load
When the checkpoint dropdown is empty, or a safetensors file throws an error when you select it, the problem is the model file or its location, not the WebUI. Confirm your checkpoints sit in the models Stable-diffusion folder of your install. Files left in Downloads or in the wrong subfolder will not appear. After adding a model, use the refresh control next to the dropdown, or restart the WebUI, since it scans the folder at startup.
If a model appears in the list but fails to load when selected, it is likely corrupt or incompletely downloaded. Compare the file size against the source; a truncated download is a classic cause of a load failure. Re download from a reliable source and try again. Extremely large or unusual model formats can also exhaust memory on load, which looks like a load failure but is really an out of memory event. For choosing models that are known good and load cleanly, the best Stable Diffusion checkpoints for NSFW roundup lists dependable options, and if your generations come out as a solid black frame after the model loads fine, that is a separate VAE or NaN issue covered in the black image fix.
Connection errored out
“Connection errored out” in the browser, or a page that spins forever trying to reach the WebUI, usually means the server is not actually up yet or crashed during startup. Check the console: if it has not printed the “Running on local URL” line, the server is still starting or died before it finished, so read the traceback for the real error. If the console shows the server is running but the browser still cannot connect, you are likely opening the wrong address or a firewall or proxy is blocking the local port. Use the exact local URL the console prints. A very long first generation can also make the UI look disconnected while the model loads into VRAM, so give the first image time before assuming a fault.
| Connection symptom | Console state | Fix |
|---|---|---|
| Browser cannot reach the page | No “Running on local URL” line | Server crashed at startup, read the traceback |
| Page loads then errors on generate | Server running, error mid job | Check VRAM and model load, read the console error |
| Spins forever on first image | Server running, model loading | Wait for the first load into VRAM to finish |
| Remote access fails | Running locally only | Launch with the listen flag, open the port |
Blank or grey UI
A blank or grey WebUI page, where the layout does not render, usually means the browser could not load the interface assets, or the server crashed after the page partially loaded. Hard refresh the browser first. If it stays blank, check the console: a crash there explains a grey page. If the console is healthy, an aggressive browser extension or strict content blocker is likely stripping the scripts the UI needs, so test in a clean incognito window. Clearing the browser cache for the local address also fixes a stale, half loaded interface after an update changed the assets.
Launch flags that fix common problems
Many Stable Diffusion failures are cured not by reinstalling but by adding the right command line argument to the set line in webui-user.bat. Knowing the handful that matter saves a lot of trial and error.
| Flag | What it fixes | When to use it |
|---|---|---|
| medvram | Out of memory on mid range cards | 6 to 8 GB VRAM struggling at higher resolution |
| lowvram | Out of memory on small cards | 4 GB cards, at the cost of speed |
| no-half-vae | Black or corrupted images from VAE | Solid black output after the model loads fine |
| skip-torch-cuda-test | Bypasses a false GPU check | Torch works but the startup test wrongly fails |
| xformers | Enables memory efficient attention | Only if a matching xformers is installed |
| reinstall-torch | Forces a fresh Torch install | After a CPU only Torch got in by mistake |
Add one flag at a time and relaunch, so you know which one helped. Do not stack lowvram and medvram together, since they conflict; pick the one that matches your card. If a flag makes things worse, remove it and relaunch. The no-half-vae flag in particular resolves a large share of the all black image complaints without any other change, and skip-torch-cuda-test rescues setups where Torch is actually fine but the launch time GPU check misfires after a driver update.
After an update broke a working install
If Stable Diffusion worked yesterday and fails today after you pulled an update or updated an extension, the update is the prime suspect. An extension that has not kept pace with the WebUI is the most frequent offender, since a single incompatible extension can crash the entire launch. Start the WebUI with extensions disabled, and if it launches cleanly, re enable them one at a time until the broken one reveals itself, then update or remove it. If the core update itself broke things, the console traceback will name the failing component, and rolling that piece back is faster than reinstalling. Treat every update as reversible, and never update the WebUI and every extension at once right before you need to generate.
If a local Stable Diffusion setup keeps failing
Local Stable Diffusion rewards patience, but not everyone wants to debug python environments and Torch builds just to generate an image. If your install keeps failing across the fixes above, a hosted no install NSFW generator gets you generating immediately, with no venv, no CUDA, and no model downloads to manage. A service like AI Nudez runs entirely server side, so you can produce results now and return to the local setup when you have time to sort it out. Plenty of people run both: the hosted tool for speed and convenience, and a local install for full control once it is stable.
A clean reinstall, done right
If the environment is too tangled to fix piecemeal, a clean reinstall is often faster, and it is less drastic than it sounds because your models and outputs are separate from the code. Back up your models Stable-diffusion folder, any LoRAs, and your generated images. Then delete the venv (or the whole install) and set it up fresh with the supported python. On the first launch let it rebuild the environment fully without interrupting it, since an interrupted first run is what breaks many installs in the first place. Add extensions back one at a time and launch after each, so a bad extension reveals itself immediately.
Before you reinstall, though, make sure you are not on a different front end that has its own fixes. If you actually run the node graph app or the Forge fork, use their dedicated guides instead of this one. And for a whole ecosystem view of diagnosing generation problems beyond launch and into output quality, the NSFW AI troubleshooting guide ties every symptom back to a cause and a fix. For choosing the right front end in the first place, the ComfyUI versus Forge versus A1111 comparison and the Forge NSFW setup guide are the natural next reads.
Frequently asked questions
Why does webui-user.bat close instantly when I launch Stable Diffusion?
The launch crashed before the UI loaded and the error scrolled past. Run the batch file from an already open command prompt, or add a pause line, so you can read the message. The most common causes are the wrong python version (stick to the supported 3.10 series), a broken venv from an interrupted first run, or antivirus quarantining a file. Delete the venv and relaunch to rebuild cleanly.
How do I fix Torch is not able to use GPU in Stable Diffusion?
A CPU only Torch build got installed, so it cannot reach your GPU. Delete Torch from the WebUI’s venv and reinstall the CUDA enabled build that matches your setup, then confirm your NVIDIA driver is current. Deleting the whole venv and relaunching forces a fresh install. If you have an AMD card, standard CUDA Torch will never work, so follow the dedicated AMD GPU guide instead.
What does Torch not compiled with CUDA enabled mean?
It means the Torch in your environment is the CPU only build, which has no CUDA support, so it cannot use your NVIDIA GPU. Uninstall that Torch from the WebUI’s virtual environment and install the CUDA build instead, or delete the venv entirely and relaunch so the WebUI reinstalls the correct GPU enabled Torch. Make sure any manual pip command targets the CUDA wheel, not the default CPU package.
Why do I get No module named errors in Stable Diffusion?
A python dependency is missing from the WebUI’s environment, usually because the venv did not finish installing or something was installed into system python instead. Delete the venv folder and relaunch so the launcher rebuilds it, or activate the WebUI’s own venv and reinstall requirements there. Confirm you are on the supported python series, since a mismatched python is the top reason a specific package refuses to install.
Why won’t my model or safetensors file load in Automatic1111?
Either the file is in the wrong place or it is corrupt. Checkpoints must sit in the models Stable-diffusion folder; after adding one, use the refresh control or restart, since the folder is scanned at startup. If a model appears in the list but fails on selection, compare its file size to the source, because a truncated download is the classic cause. Re download from a reliable source and try again.
How do I fix connection errored out in Stable Diffusion?
Check the console for the Running on local URL line. If it is missing, the server crashed at startup, so read the traceback for the real error. If the server is running but the browser cannot connect, you are opening the wrong address or a firewall is blocking the local port, so use the exact URL the console prints. A slow first generation can also look like a disconnect while the model loads into VRAM.
Do xformers errors stop Stable Diffusion from working?
They can, because an xformers build must match your Torch and python versions, and a mismatch crashes the launch. Remove the xformers flag to launch without it and confirm that is the culprit. If the WebUI then starts, reinstall a matching xformers version or simply run without it, since modern Torch attention is fast enough that xformers is optional. Never mix an xformers build with a different Torch version.
Should I use ComfyUI or Forge fixes instead of these?
Yes, if you actually run those front ends. This guide targets the general Stable Diffusion and Automatic1111 layer. The node graph app has its own failure modes like red missing nodes and the Reconnecting websocket loop, covered in the ComfyUI guide, while the Forge fork has update and extension incompatibilities covered in the Forge guide. Using the right front end guide saves you from applying an Automatic1111 fix that does not match your app.
