VisoMaster Fusion: a Windows desktop face swapper with a portable launcher

What VisoMaster Fusion actually is, and who it is for

VisoMaster Fusion is a desktop application for face swapping, enhancement and editing on images, videos, live webcam feeds and VR180 footage. The README describes it as building on the original VisoMaster work by @argenspin and @Alucard24 with contributions from the wider community, so this is a fork-lineage project rather than a from-scratch codebase.

The audience is narrower than the feature list suggests. The stated system requirements are Windows 10 or Windows 11, 64-bit, with an Nvidia GPU recommended, 6 GB of VRAM minimum and 8-12 GB recommended for heavier work. CPU execution is possible, and the README says AI processing is much slower that way. If you are on macOS or Linux, the repository does contain Start_mac.sh and requirements_mac.txt, but the README's system requirements section names only Windows, so a Mac user is working outside the documented path.

The practical user is someone editing video who wants a graphical workflow rather than a script: source face cards, per-frame timeline markers, a job queue, and output controls for multi-segment renders. That is a different product shape from a command-line swapper, and the difference matters when you decide what to install.

The swapper pipeline: models, masks and the second restorer pass

The mechanism visible in the README is a staged pipeline. A face detector finds faces in the frame, a swapper model replaces them, masks decide which pixels are affected, and restorers or enhancers clean up the result.

Swapper models listed include Inswapper128, InStyleSwapper variants, SimSwap, GhostFace, CSCS, and DeepFaceLive DFM models. That is a wide menu, and the README does not rank them, so model choice is left to experimentation. Multi-face work uses source face cards, saved embeddings, similarity thresholding, and optional ByteTrack tracking, which is how the app keeps a consistent identity attached to the right person across frames.

Masking is where most of the visible quality lives. The README lists occlusion, XSeg, text, face parser, border, profile-angle, and mouth-focused mask controls. Restoration is similarly layered: face restorers, a second restorer pass, auto restore blending, GFPGAN-1024 support, and frame enhancers. There is also a ReF-LDM Denoiser with single-step and DDIM modes that the README says can be applied at multiple pipeline points.

Inference runs through CUDA, TensorRT, TensorRT-Engine, or CPU providers, with TensorRT as the default. TensorRT is the fastest path in most Nvidia setups, but it also means engine compilation and model optimization steps that the launcher exposes as separate tools. The README does not publish timings for any provider, so treat the default as a starting point rather than a measured recommendation.

How to install VisoMaster Fusion with the portable launcher

The README states that most users should use the portable launcher and that the first launch downloads the portable runtime, dependencies, FFmpeg, and model files. Create a new folder, download only Start_Portable.bat from the latest release, place it in that folder, and run it. After setup, the README says to always start the application with that same file.

# Download Start_Portable.bat from the latest release page,
# place it in a new empty folder, then run it.
Start_Portable.bat

The launcher installs everything into that folder, including Python 3.12, Git, FFmpeg, PyTorch, CUDA Toolkit, TensorRT, cuDNN, ONNX Runtime GPU, and the required model files. You do not need the non-portable steps for this path. Internet access is required on first run, and the README recommends 20-30 GB of free disk space.

There is a warning worth repeating: the project says it is only distributed through this repository, and that sites like visomaster.com and visomaster.org are not affiliated with the maintainers. Downloading a build from anywhere else is a supply-chain risk the maintainers have explicitly flagged.

If you want to manage the environment yourself, the non-portable route starts by cloning the repository, creating a Python 3.12 environment, and installing requirements_cu13.txt with uv. The README also notes that the main branch is the normal choice and the dev branch carries newer or in-progress changes.

A manual install for a first render

The non-portable path is documented for people who want their own Python environment. Clone the repository and create the environment first.

git clone https://github.com/VisoMasterFusion/VisoMaster-Fusion
cd VisoMaster-Fusion
conda create -n visomaster python=3.12 -y
conda activate visomaster
pip install uv

Install the pinned requirements and pull the model files. The download script is the documented way to get the models the app expects.

uv pip install -r requirements_cu13.txt
python download_models.py

FFmpeg is a separate dependency. On Windows the README gives a winget command and a manual build.

winget install -e --id Gyan.FFmpeg --version 7.1.1

Then start the app. Either open Start.bat, or run main.py from inside the VisoMaster-Fusion directory with the environment active.

python main.py

What you should see is the desktop UI, from which you load a source face image and a target video or webcam feed, pick a swapper model, apply masks and a restorer, and render. The README does not document first-run defaults for model selection, so the first render is a calibration exercise: start at low resolution, check the mask edges, then raise quality once the identity holds.

Where VisoMaster Fusion breaks down

The largest constraint is platform. The system requirements section names Windows 10 and Windows 11 only. A Start_mac.sh file exists in the repository root and requirements_mac.txt is present, but the README does not state supported macOS versions, chip requirements, or whether the GPU providers behave the same way. Anyone searching for the best face swapping software for Mac should treat the Mac path as undocumented rather than supported.

Second, the portable launcher is a single point of failure by design. It downloads a runtime, dependencies, FFmpeg, and models on first run, and the README says to always start the app through it. The launcher does expose repair, dependency check, model check, model restore, and version rollback tools, which is more recovery surface than many projects offer, but the README does not document rollback behaviour in detail, so the exact recovery path after a failed update is not something you can read up on in advance.

Third, the model menu is wide and unranked. Inswapper128, InStyleSwapper variants, SimSwap, GhostFace, CSCS, and DFM models all appear in the same list with no guidance on when each is appropriate. That is a real cost for a new user, and it is the kind of decision the README leaves to the community.

Finally, VRAM. The README gives 6 GB as a minimum for basic use and 8-12 GB or more for heavier workflows. A second restorer pass, GFPGAN-1024, frame enhancers, and the ReF-LDM Denoiser all compete for the same memory. If you are on a 6 GB card, expect to choose between features rather than stack them.

VisoMaster Fusion compared with FaceFusion

FaceFusion is the comparison people search for, and the difference in approach is architectural rather than cosmetic. FaceFusion is built around a command-line core with a separate web interface, which makes it scriptable and straightforward to run headless on a server or inside a batch pipeline. VisoMaster Fusion is a desktop application with a Qt-style GUI, a Job Manager for saving workspaces and running batches unattended, and a portable launcher that manages the whole runtime for you.

That trade runs in both directions. VisoMaster Fusion gives you timeline markers for per-frame settings, issue scanning and dropped-frame review tools before output, and a layered mask and restorer stack exposed as GUI controls. Those are conveniences you feel immediately when editing a difficult clip. What you give up is the ability to drop the tool into a shell script or a container without the GUI, and the ability to run it on a Linux box as a documented configuration.

The two projects also differ in how much they ask you to know. FaceFusion's CLI model pushes configuration into flags and files. VisoMaster Fusion pushes it into the interface, which is friendlier until you want to reproduce a render exactly or move it to another machine. The Job Manager partly answers that, since it saves workspaces and loads jobs, but the README does not describe the job file format, and pyproject.toml only mentions serialization tests in the test-suite description.

Maintenance, updates and the GPL-3.0 licence

The repository is not archived, and the last push was on 2026-09-06. Releases are frequent: v3.12.3 on 2026-08-29, v3.9.3 on 2026-07-27, and v3.8.0 on 2026-07-14. The README states that future releases use PR label-based versioning based on Semantic Versioning.

Upgrade cost depends on which install you chose. The portable launcher bundles update, repair, dependency check, model check, model optimization, model restore, launcher self-update, and version rollback tools, so updates are handled inside the launcher. The non-portable path is the usual three commands: git pull, reinstall requirements_cu13.txt, and re-run download_models.py. Note that requirements_cu13.txt targets CUDA 13, and pyproject.toml points uv at pypi.nvidia.com, the PyTorch cu130 index, an ONNX Runtime CUDA 13 nightly index, and pypi.org with index-strategy set to unsafe-best-match. Nightly indexes move, which is a plausible source of install breakage between releases.

The licence is GPL-3.0. For most individual users that changes nothing. If you plan to distribute a modified build, or to bundle the application inside a commercial product, the copyleft terms apply to the combined work, and that is a question for a lawyer rather than a review. The README's own distribution warning is separate from licensing: it says the project is only distributed through this repository and that visomaster.com and visomaster.org are not affiliated.

Editorial conclusion

Adopt VisoMaster Fusion if you are on Windows 10 or 11 with an Nvidia GPU and 8-12 GB of VRAM, and you want a GUI for multi-face video swaps with batch jobs and model controls. Skip it if you need macOS, Linux, or a headless pipeline; the README points macOS users at Start_mac.sh and requirements_mac.txt but documents no macOS requirements, and the stated system requirements are Windows only. Verify two things before committing: that the portable launcher runs in a fresh folder on your machine, and that your chosen swapper model and restorer fit your VRAM budget at the resolution you intend to render.

Frequently asked questions

Official sources

1. Issues
2. License: GPL-3.0
3. README
4. Releases
5. VisoMasterFusion/VisoMaster-Fusion on GitHub