VisoMaster RTX 5090 CUDA Fix: sm_120, Drivers & Setup

If the error explicitly mentions sm_120 or no kernel image is available, check the runtime in the VisoMaster installation you actually launched. A newer NVIDIA driver alone does not replace an old PyTorch or ONNX Runtime inside an archive. Conversely, an updated archive still needs a compatible driver.

Documentation reviewed October 5, 2026. Package filenames and sizes below were checked on September 23, 2026. This is a diagnosis and installation guide; we have not run a controlled RTX 5090 benchmark for this article.

Match the error before choosing a fix

What you see	First thing to check	Next step
sm_120 is not compatible	GPU architecture support in the packaged runtime	Use the archive intended for RTX 50 Series, or the source project's supported dependency set
no kernel image is available	The first failing library in the complete log	Check its GPU support; this error is not exclusive to RTX 50 GPUs
CUDA provider fails to load	NVIDIA driver and the archive's runtime libraries	Follow the matching package instructions; avoid mixing libraries from two installations
CUDA out of memory	Resolution, enhancement and other GPU applications	Close unnecessary GPU tasks; retry a short, smaller clip before increasing settings
Window closes immediately	Extraction, folder path and full console error	Run the supplied launcher from a terminal to retain its error message
Browser cannot open port 7860	Whether the launcher is still running and the port is occupied	Read the console; use the README's port setting if needed

Why RTX 50 Series can expose an old-runtime problem

The RTX 50 family uses NVIDIA's Blackwell architecture. PyTorch's official 2.7 release introduced Blackwell support and CUDA 12.8 wheels. That provides context for the error, not proof that replacing one library will fix every VisoMaster distribution: an application may also depend on ONNX Runtime, TensorRT and version-specific model components.

Use the dependency versions supported by your chosen build. Do not paste a generic PyTorch upgrade into a working all-in-one installation.

Choose and install the prepared Windows archive

• VisoMaster_50Series.zip: 27.86 GB compressed, directory timestamp 2026-03-29. The included README identifies this as the RTX 50 Series package.
• VisoMaster-v0.1.6.zip: 16.06 GB compressed, directory timestamp 2025-10-14. Check its own instructions for older hardware.

Timestamps are not release guarantees. You need space for both the archive and extraction, plus models and exports; 20 GB free space is insufficient for the RTX 50 archive itself.

1. Open the download selection page and review the current Patreon post's access conditions and archive instructions.
2. Confirm your GPU and update its driver if required by that build. Download the matching archive and README from the post's provided folder.
3. Extract to a new short English-only path, for example D:\VisoMaster_50Series. Keep the old working installation separate.
4. Start VisoMaster.exe or Run.bat as directed by that archive. Keep the launcher open and visit http://localhost:7860.
5. Test one face on a short clip first. Confirm preview, completed export and audio sync before a long job.

This is a paid prepared package. The original source project remains available separately.

Review the Windows package on Patreon

If it still fails

Record the exact archive name, application version, Windows version, GPU, driver, first error and the action that triggered it. Keep the complete local log, but remove personal paths before sharing it. A memory error, missing model and architecture error need different fixes; downloading the same archive repeatedly will not distinguish them.

If you installed from source, use the original project's installation instructions. Do not combine those commands with a packaged runtime unless its README specifically directs you to do so.

Common questions

Does this guarantee every RTX 30, 40 and 50 card will work?

No. Check the selected archive's compatibility information. A GPU family label does not establish tested driver, VRAM and runtime combinations for every desktop or laptop.

Do I need the CUDA toolkit?

A prepared archive can include runtime dependencies, while source installations have their own requirements. A compatible NVIDIA driver is still required. Follow your build's instructions rather than installing extra toolkits indiscriminately.

Is the RTX archive the latest VisoMaster Fusion?

The filename does not prove that. Check its README and version for the Fusion features you need before purchasing.

Related guides and sources

• System requirements
• VisoMaster vs FaceFusion: setup and workflow
• Face swap tutorial
• PyTorch 2.7: Blackwell support