Fix NVIDIA Driver Performance Issues in Stable Diffusion WebUI

How to Fix It

Solution 1: Upgrade to the Latest NVIDIA Driver (Recommended)

NVIDIA has resolved this issue starting with driver version 537.58 (537.42 for Studio drivers). As of October 31, 2023, the GitHub issue maintainers confirmed that the problem should be entirely resolved with driver version 546.01 or newer. Follow these steps:

1. Check your current driver version: Open the NVIDIA Control Panel (right-click on desktop) and go to Help > System Information. Look for the "Driver Version" field.
2. Download the latest driver: Go to the official NVIDIA driver download page at https://www.nvidia.com/download/Find.aspx. Select your GPU model and operating system, then download the latest Game Ready or Studio driver (version 546.01 or newer).
3. Perform a clean installation: When running the installer, select "Custom (Advanced)" and check the box for "Perform a clean installation." This removes all previous driver files and settings, reducing the chance of conflicts.
4. Restart your computer: After installation, reboot to ensure the new driver is fully loaded.
5. Test Stable Diffusion WebUI: Launch the WebUI and run a few generations. Performance should return to normal.

Expected outcome: Generation speeds should be restored, and no more stalling or slowdowns during heavy memory usage.

Solution 2: Disable System Memory Fallback (For Driver 546.01+)

If you are on driver version 546.01 or newer and still experience issues, NVIDIA has published a help article to disable the system memory fallback behavior. Follow the guide at https://nvidia.custhelp.com/app/answers/detail/a_id/5490. The exact steps from that article are:

1. Open the NVIDIA Control Panel.
2. Navigate to "Manage 3D Settings."
3. Under "Global Settings," find the option "System Memory Fallback Policy" (or similar, depending on driver version).
4. Set it to "Off" or "Disabled."
5. Apply the changes and restart Stable Diffusion WebUI.

Note: This setting may not appear in older drivers. If you cannot find it, upgrade to the latest driver first.

Solution 3: Downgrade to Driver Version 531 or Below (Fallback)

If upgrading does not work or you need an immediate fix, downgrading to a driver version before the memory management change is a reliable workaround. This was the original recommendation from the GitHub issue discussion (Source 1).

1. Uninstall your current driver: Use Display Driver Uninstaller (DDU) in Safe Mode to completely remove the existing NVIDIA driver. This prevents conflicts between old and new driver files.
2. Download driver version 531 or below: Go to https://www.nvidia.com/download/Find.aspx and manually select a driver version from the list. Look for versions like 531.79 or earlier. You may need to use third-party archives like Guru3D's driver collection if NVIDIA no longer hosts older versions.
3. Install the older driver: Run the installer with a clean installation as described in Solution 1.
4. Restart and test: Reboot and verify that Stable Diffusion WebUI runs without slowdowns.

Caveat: Older drivers may lack security updates and optimizations for newer games or applications. Only use this as a temporary measure until you can upgrade to a fixed driver.

Solution 4: Community-Reported Workarounds

Some users on the GitHub issue thread reported that the issue persisted even after driver updates. For these cases, the following community-sourced workarounds may help:

• Reduce batch size or image resolution: Lowering the memory footprint can avoid triggering the fallback behavior. Try generating one image at a time instead of batches, or reduce the resolution to 512x512.
• Use the --medvram or --lowvram flags: When launching Stable Diffusion WebUI, add these flags to the command line to optimize memory usage. For example:

  python launch.py --medvram
  

  or

  python launch.py --lowvram
  

• Monitor VRAM usage: Use tools like GPU-Z or the Task Manager to check if VRAM is being exhausted. If it is, the fallback behavior may be triggered.

If Nothing Works

If none of the above solutions restore performance, consider the following escalation paths:

• Check the NVIDIA driver release notes: Visit https://us.download.nvidia.com/Windows/ and look for the latest release notes PDF. Search for bug number 4172676 to see if a fix is listed.
• File a bug report with NVIDIA: Use the NVIDIA Developer Bug Report tool or contact NVIDIA support directly. Reference the tracking number 4172676.
• Post on the Stable Diffusion WebUI GitHub issue: The original issue at https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/11063 is still open for discussion. Provide your driver version, GPU model, and a detailed description of the problem.
• Use an alternative WebUI fork: Some users have reported better stability with forks like Vladmandic's automatic (https://github.com/vladmandic/automatic), which may have additional memory management optimizations.

How to Prevent It

To avoid this issue in the future:

• Keep your NVIDIA drivers updated: Regularly check for new driver releases, especially if you use Stable Diffusion frequently. NVIDIA has acknowledged this bug and is actively working on fixes.
• Monitor driver release notes: Before updating, read the release notes PDF for any known issues related to Stable Diffusion or creative applications. Look for mentions of bug 4172676.
• Use a stable driver version: If you find a driver version that works well (e.g., 531 or below), consider sticking with it until you have confirmed that a newer version does not introduce regressions.
• Limit VRAM usage: Configure Stable Diffusion WebUI to use less VRAM by default. Use the --medvram flag in your launch script to reduce the chance of hitting the fallback behavior.
• Back up your current driver: Before updating, note your current driver version so you can easily downgrade if needed.