First Windows output: evidence status and next test

Evidence reviewed and page updated: 7 October 2026. The earlier Windows screenshots record installation and Gradio startup. The screenshot dated 21 September 2026 shows loaded inputs with an empty result panel. No generated Windows MP4, exact repository revision, complete environment manifest or full-decode record is currently attached to this page. Native Windows first-output verification is pending; startup alone is not an end-to-end inference pass. This is an unverified route, not a recorded new failure.

Capture a real first output on the existing Windows machine

Download the Windows first-output capture kit. It includes the shared permitted MP4/WAV and a capture helper. The helper is prepared and syntax checked, but has not yet been executed on Windows. It does not install or change dependencies. Use native Windows with the existing NVIDIA GPU environment; WSL and SSH into a Linux rental must be identified as separate routes.

1. Extract the kit. Open Miniconda Prompt and activate the actual MuseTalk environment, for example conda activate MuseTalk.
2. Run capture-first-output.cmd from that prompt and enter the full MuseTalk repository folder. The helper records the exact command, Git revision, GPU/driver, Python, library versions and elapsed time.
3. Keep the resulting windows-evidence-*.zip whether generation succeeds or fails. A pass requires a newly created MP4, video and audio streams and a complete FFmpeg decode.
4. Open and watch/listen to the saved windows-result.mp4. Capture the real final terminal output and the Gradio result if using the interface. Preserve original screenshots; add only annotations on copies.
5. If it fails, preserve the entire error and environment, apply one targeted repair at a time and run the capture again. Do not call an import check or a Gradio URL a completed video test.

Repair notes and evidence limits

The original guide below records proposed or previously reported repairs. They have not been retested on native Windows in this 7 October review. Match the actual traceback before applying any change:

• FFmpeg path or command not found: check where ffmpeg, where ffprobe and the actual folder containing both executables. The guide’s set PATH=... command is for CMD/Miniconda Prompt. Use the shell-specific FFmpeg instructions for other shells.
• MMPose/Chumpy import or install errors: retain the complete error and package versions before following the corresponding repair section. After the repair, require a successful import and a new inference output, not just an installer exit code.
• Hugging Face dependency conflict: record the actual installed versions and the constraint printed in the error; do not treat the older pin below as a universal repair for unrelated environments.
• Missing model files: confirm filenames, paths and the model version using the model-file guide. An empty result panel by itself does not identify the cause.

For an independently verified fallback, the Linux video tutorial provides real inputs, output and environment records. Its GPU output is explicitly a Linux result and does not establish native Windows success.

This guide documents the Windows installation steps and earlier recorded Gradio startup on an NVIDIA GPU. Use the first-output evidence checklist below before treating the installation as an end-to-end generation pass.

I tested this installation directly on Windows with Python 3.10, PyTorch 2.0.1 + CUDA 11.8, MuseTalk 1.5, FFmpeg, and an NVIDIA GPU. The test environment successfully reached the MuseTalk Gradio server at:

http://127.0.0.1:7860

The official MuseTalk repository recommends Python 3.10 and provides Windows instructions for PyTorch, FFmpeg, model downloads, and the Gradio application.

This guide also includes fixes for the Windows errors I encountered, including:

• conda setup
• FFmpeg not recognized
• incorrect FFmpeg path
• missing MuseTalk model files
• huggingface_hub version conflict
• ModuleNotFoundError: No module named 'mmpose'
• Chumpy installation failure
• CUDA verification
• starting the MuseTalk Gradio interface

MuseTalk Windows Requirements

Before installing MuseTalk, you should have:

• Windows 10, Windows 11, or Windows Server
• 64-bit operating system
• NVIDIA GPU
• Current NVIDIA GPU driver
• Internet connection
• Around 15–20 GB or more of free disk space
• Miniconda
• Git
• FFmpeg

The upstream Gradio documentation reports a Windows test using an RTX 3050 Ti Laptop GPU with 4 GB VRAM. This is upstream evidence, not a hardware minimum measured by this site’s first-output test.

Step 1: Check Your NVIDIA GPU

Open Command Prompt and run:

nvidia-smi

You should see your NVIDIA GPU, driver version, available VRAM, and CUDA driver information.

If Windows says:

'nvidia-smi' is not recognized

install or update the NVIDIA driver first.

You do not need to install the full CUDA Toolkit just because the nvidia-smi output displays a different CUDA version. We will install the CUDA-enabled PyTorch build required by this MuseTalk setup separately.

---

Step 2: Install Miniconda on Windows

MuseTalk is much easier to manage inside a separate Conda environment. This prevents its Python packages from conflicting with other projects installed on your computer.

Download the Windows Miniconda installer from the official Conda/Anaconda source and run the .exe installer.

The official Conda documentation recommends running the installer, following the setup wizard, and then opening the Miniconda/Anaconda Command Prompt from the Start menu.

You can normally keep the default installation options.

After Miniconda is installed, open:

Miniconda Prompt

or:

Anaconda Prompt

Then check that Conda works:

conda --version

For example, my test returned:

conda 26.7.1

If Conda Asks You to Accept the Terms of Service

Recent Miniconda releases can display an Anaconda Terms of Service prompt when you first access the default package channels.

Read the prompt and accept it if you agree to those terms. Anaconda introduced this command-line ToS mechanism for access to its repositories.

After acceptance, rerun the Conda command.

---

Step 3: Create a Python 3.10 MuseTalk Environment

Create a separate environment named MuseTalk:

conda create -n MuseTalk python=3.10 -y

MuseTalk currently recommends Python 3.10.

In my Windows test, Conda created the environment under:

C:\Users\Administrator\miniconda3\envs\MuseTalk

and installed Python 3.10 successfully.

Now activate it:

conda activate MuseTalk

Your command prompt should change from something similar to:

(MuseTalk) C:\Users\Administrator>

Check Python:

python --version

You should get:

Python 3.10.x

My installation used Python 3.10.21.

Conda Command Not Recognized?

If:

conda --version

returns:

'conda' is not recognized as an internal or external command

the easiest fix is to use Miniconda Prompt instead of a standard Command Prompt.

Close your existing CMD window, search Windows for:

Miniconda Prompt

and continue the installation there.

---

Step 4: Install Git

Check whether Git is already installed:

git --version

If you get a Git version, continue.

If Windows says Git is not recognized, install Git for Windows and reopen your terminal.

Now clone the official MuseTalk repository:

git clone https://github.com/TMElyralab/MuseTalk.git

Then enter the project directory:

cd MuseTalk

Your prompt should now look similar to:

(MuseTalk) C:\Users\Administrator\MuseTalk>

My test successfully cloned the repository and downloaded approximately 25 MB of repository files before model weights were installed separately.

---

Step 5: Install PyTorch With CUDA Support

MuseTalk’s current installation instructions specify PyTorch 2.0.1 with torchvision 0.15.2 and torchaudio 2.0.2.

Run:

pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118

The Windows CUDA 11.8 PyTorch package is large. In my test, the PyTorch wheel alone was approximately 2.6 GB.

Verify That PyTorch Can See Your GPU

Run:

python -c "import torch; print('Torch:', torch.__version__); print('CUDA available:', torch.cuda.is_available()); print('CUDA:', torch.version.cuda); print('GPU:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'CPU')"

You want:

CUDA available: True

If it says:

CUDA available: False

do not continue with GPU inference yet. Check your NVIDIA driver and PyTorch installation.

---

Step 6: Install MuseTalk Python Requirements

Make sure you are still inside:

(MuseTalk) C:\Users\Administrator\MuseTalk>

Then run:

pip install -r requirements.txt

The requirements install MuseTalk’s Python libraries including Transformers, Diffusers, Gradio, OpenCV, Librosa, TensorFlow-related dependencies, FFmpeg Python bindings, and other packages.

The tested installation pulled transformers==4.39.2, gradio==5.24.0, diffusers==0.30.2, and other versions specified by the repository.

---

Step 7: Install OpenMMLab Dependencies

MuseTalk uses OpenMMLab packages for pose and face processing.

First install OpenMIM:

pip install --no-cache-dir -U openmim

Then run:

mim install mmengine

Install MMCV:

mim install "mmcv==2.0.1"

Install MMDetection:

mim install "mmdet==3.1.0"

The official MuseTalk instructions currently specify these packages and versions.

Important Windows Fix Before Installing MMPose

On my Windows installation, running:

mim install "mmpose==1.1.0"

directly failed while trying to build chumpy.

The error looked like:

ModuleNotFoundError: No module named 'pip'

ERROR: Failed to build 'chumpy' when getting requirements to build wheel

You can avoid this error by installing Chumpy first.

Run:

pip install setuptools wheel

Then:

pip install chumpy==0.70 --no-build-isolation

My Windows test successfully built and installed Chumpy using this method.

Now install MMPose:

mim install "mmpose==1.1.0"

Verify it:

python -c "import mmpose; print(mmpose.__version__)"

You should see:

1.1.0

The final tested environment successfully installed mmpose-1.1.0 along with xtcocotools, json-tricks, munkres, and Cython.

---

Step 8: Install FFmpeg on Windows

FFmpeg is required because MuseTalk needs it for reading, processing, and exporting video/audio.

There are two ways to install FFmpeg.

Option 1: Install FFmpeg With Winget

First check whether Winget is available:

winget --version

Then install FFmpeg:

winget install -e --id Gyan.FFmpeg

Gyan.FFmpeg is available in Microsoft’s Winget package repository.

Close the terminal after installation and open a new one.

Then test:

ffmpeg -version

If FFmpeg information appears, you can continue.

If:

'ffmpeg' is not recognized as an internal or external command

use the manual method below.

---

Option 2: Manually Install FFmpeg on Windows

This is the method I used during testing.

Download a Windows FFmpeg build. I used an Essentials build from Gyan.

The downloaded ZIP in my test was:

ffmpeg-9.0.2-essentials_build.zip

The exact version number can change, so yours may have a different filename.

Extract the archive.

For example, I placed it under:

C:\ffmpeg\

My final FFmpeg executable was located at:

C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin\ffmpeg.exe

Open the extracted bin directory and confirm that ffmpeg.exe exists.

Test FFmpeg Using the Full Path

Run:

C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin\ffmpeg.exe -version

Change the folder name if your FFmpeg version is different.

Important: use:

-version

not:

--version

During my test, --version returned:

Unrecognized option '-version'.
Error splitting the argument list: Option not found

while -version worked correctly.

Fix: ffmpeg Is Not Recognized

You may successfully run ffmpeg.exe with its full path but still get:

'ffmpeg' is not recognized as an internal or external command

when running:

ffmpeg -version

That happened in my Windows test as well.

Temporarily add FFmpeg to PATH:

set PATH=C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin;%PATH%

Then test again:

ffmpeg -version

It should now work. In my test, the command worked immediately after adding the FFmpeg bin directory to PATH.

This set PATH=... method applies to the current CMD session.

You can also permanently add the FFmpeg bin folder through:

Windows Search
→ Edit the system environment variables
→ Environment Variables
→ Path
→ Edit
→ New

Add:

C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin

The official MuseTalk documentation also recommends adding the FFmpeg bin directory to Windows PATH and verifying it with ffmpeg -version.

---

Step 9: Download MuseTalk Model Weights

MuseTalk includes a Windows batch file for downloading its required models.

Run this from the MuseTalk directory:

download_weights.bat

The model download is several gigabytes, so give it time to finish.

The script downloads MuseTalk, SD VAE, Whisper, DWPose, SyncNet, and Face Parse weights. The current official batch file downloads them into the models directory.

After it finishes, you may see:

All weights have been downloaded successfully!

Important: Fix Hugging Face Hub After Running the Script

There is an important issue I encountered with the current Windows weight script.

The batch file runs:

pip install -U "huggingface_hub[hf_xet]"

which can upgrade huggingface_hub to a 1.x release. The current batch file contains this upgrade command.

In my test it installed:

huggingface_hub 1.32.0

but transformers 4.39.2 required:

huggingface_hub < 1.0

That produced a dependency conflict.

After the model download finishes, run:

pip uninstall huggingface_hub -y

Then install the tested compatible version:

pip install huggingface_hub==0.36.2

Confirm:

python -c "import huggingface_hub; print(huggingface_hub.__version__)"

Expected:

0.36.2

---

Step 10: Check for Missing Model Files

Do not assume that:

All weights have been downloaded successfully!

means every required file actually exists.

This happened during my test.

MuseTalk started and reported:

The following required model files are missing:
- SD VAE
- Whisper
- Face Parse

The current MuseTalk application specifically checks for:

models\sd-vae\config.json
models\whisper\config.json
models\face-parse-bisent\79999_iter.pth

along with its other required weights.

Check all three automatically:

python -c "import os; files=['models/sd-vae/config.json','models/whisper/config.json','models/face-parse-bisent/79999_iter.pth']; [print(('OK      ' if os.path.exists(f) else 'MISSING ')+f) for f in files]"

You want:

OK      models/sd-vae/config.json
OK      models/whisper/config.json
OK      models/face-parse-bisent/79999_iter.pth

Fix Missing SD VAE Config

Run:

hf download stabilityai/sd-vae-ft-mse config.json --local-dir models\sd-vae

Fix Missing Whisper Config

Run:

hf download openai/whisper-tiny config.json --local-dir models\whisper

Fix Missing Face Parse Model

Run:

hf download ManyOtherFunctions/face-parse-bisent 79999_iter.pth --local-dir models\face-parse-bisent

These commands fixed all three missing files in my test.

The completed folders looked like:

models
├── musetalk
├── musetalkV15
├── syncnet
├── dwpose
├── face-parse-bisent
│   ├── 79999_iter.pth
│   └── resnet18-5c106cde.pth
├── sd-vae
│   ├── config.json
│   └── diffusion_pytorch_model.bin
└── whisper
    ├── config.json
    ├── preprocessor_config.json
    └── pytorch_model.bin

This matches the model layout currently documented by MuseTalk.

---

Step 11: Run MuseTalk Gradio on Windows

Now you can start MuseTalk.

Make sure the Conda environment is active:

conda activate MuseTalk

Enter the repository if necessary:

cd C:\Users\Administrator\MuseTalk

Change that path if you cloned MuseTalk somewhere else.

Then start the Gradio interface:

python app.py --use_float16 --ffmpeg_path C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin

Replace the FFmpeg folder with your actual location.

The --use_float16 option reduces VRAM requirements and is the mode recommended in the official example for Windows.

On the first start, MuseTalk may download an additional face detection checkpoint.

In my test I saw:

cuda start

followed by:

load unet model from ./models/musetalkV15/unet.pth

and finally:

* Running on local URL: http://127.0.0.1:7860

That means MuseTalk is running.

Open your browser and visit:

http://127.0.0.1:7860

---

Step 12: Use the MuseTalk Gradio Interface

Once the web interface opens, provide your source face/video and audio input through the MuseTalk interface.

For your first test, use a short clip and short audio file. This makes it easier to confirm that CUDA, face detection, audio processing, FFmpeg, and video export are all working.

MuseTalk also allows you to generate the first frame while adjusting lip-sync-related parameters before processing the complete output. The project recommends this approach to help tune the result before spending time on the complete video.

---

Warnings You May See During Startup

diffusion_pytorch_model.safetensors Not Found

My successful Windows startup displayed:

An error occurred while trying to fetch models\sd-vae:
Error no file named diffusion_pytorch_model.safetensors found

followed by:

Defaulting to unsafe serialization.

The application then continued loading unet.pth and successfully started the Gradio server at:

http://127.0.0.1:7860

So if you see this message but the local Gradio URL appears afterwards, MuseTalk has still started.

---

Fix: ModuleNotFoundError: No module named 'mmpose'

If you get:

ModuleNotFoundError: No module named 'mmpose'

run:

pip install setuptools wheel

Then:

pip install chumpy==0.70 --no-build-isolation

Then:

mim install "mmpose==1.1.0"

Verify:

python -c "import mmpose; print(mmpose.__version__)"

Then run MuseTalk again.

---

Fix: Chumpy No module named 'pip'

If MMPose installation produces:

ERROR: Failed to build 'chumpy'

or:

ModuleNotFoundError: No module named 'pip'

do not repeatedly run the MMPose installation.

Install Chumpy separately:

pip install setuptools wheel
pip install chumpy==0.70 --no-build-isolation

Then retry:

mim install "mmpose==1.1.0"

This was the fix that worked in my Windows installation.

---

Fix: Hugging Face Version Conflict

If you get something similar to:

ImportError: huggingface-hub>=0.19.3,<1.0 is required

check the version:

python -c "import huggingface_hub; print(huggingface_hub.__version__)"

If it shows 1.x, run:

pip uninstall huggingface_hub -y
pip install huggingface_hub==0.36.2

Then start MuseTalk again.

---

Fix: Required Model Files Are Missing

If MuseTalk displays:

The following required model files are missing:
- SD VAE
- Whisper
- Face Parse

run:

hf download stabilityai/sd-vae-ft-mse config.json --local-dir models\sd-vae

hf download openai/whisper-tiny config.json --local-dir models\whisper

hf download ManyOtherFunctions/face-parse-bisent 79999_iter.pth --local-dir models\face-parse-bisent

Then retry:

python app.py --use_float16 --ffmpeg_path C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin

---

Fix: FFmpeg Path Not Found

This will fail if your actual installation is:

C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin

but you run:

C:\ffmpeg\bin\ffmpeg.exe -version

Windows will report:

The system cannot find the path specified.

That exact path mistake occurred during my test.

Find the folder containing:

ffmpeg.exe

and pass that bin directory to MuseTalk.

For example:

python app.py --use_float16 --ffmpeg_path C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin

MuseTalk’s application can add the supplied FFmpeg directory to PATH automatically if FFmpeg is not already accessible.

---

Check MuseTalk GPU Usage

While generating a video, open another Command Prompt window and run:

nvidia-smi -l 1

This refreshes NVIDIA GPU statistics every second.

Look at:

GPU-Util
Memory-Usage

Press:

Ctrl + C

to stop monitoring.

---

Starting MuseTalk Again Later

You do not need to repeat the installation every time.

Open Miniconda Prompt and run:

conda activate MuseTalk

Go to MuseTalk:

cd C:\Users\Administrator\MuseTalk

Then start it:

python app.py --use_float16 --ffmpeg_path C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin

Open:

http://127.0.0.1:7860

That’s all you need for future sessions.

---

Complete MuseTalk Windows Command List

Here is the main installation sequence in one place:

conda create -n MuseTalk python=3.10 -y

conda activate MuseTalk

python --version

git clone https://github.com/TMElyralab/MuseTalk.git

cd MuseTalk

pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118

pip install -r requirements.txt

pip install --no-cache-dir -U openmim

mim install mmengine

mim install "mmcv==2.0.1"

mim install "mmdet==3.1.0"

pip install setuptools wheel

pip install chumpy==0.70 --no-build-isolation

mim install "mmpose==1.1.0"

download_weights.bat

pip uninstall huggingface_hub -y

pip install huggingface_hub==0.36.2

python -c "import huggingface_hub; print(huggingface_hub.__version__)"

python -c "import os; files=['models/sd-vae/config.json','models/whisper/config.json','models/face-parse-bisent/79999_iter.pth']; [print(('OK      ' if os.path.exists(f) else 'MISSING ')+f) for f in files]"

If those three model files are missing:

hf download stabilityai/sd-vae-ft-mse config.json --local-dir models\sd-vae

hf download openai/whisper-tiny config.json --local-dir models\whisper

hf download ManyOtherFunctions/face-parse-bisent 79999_iter.pth --local-dir models\face-parse-bisent

For a manually installed FFmpeg:

set PATH=C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin;%PATH%

ffmpeg -version

Finally:

python app.py --use_float16 --ffmpeg_path C:\ffmpeg\ffmpeg-9.0.2-essentials_build\bin

Then open:

http://127.0.0.1:7860

Frequently Asked Questions

1. Does MuseTalk work on Windows?

Yes. MuseTalk’s official repository provides Windows-specific model download, FFmpeg, inference, and Gradio instructions.

2. Which Python version should I use for MuseTalk?

Use Python 3.10 for this setup.

conda create -n MuseTalk python=3.10 -y

3. Do I need an NVIDIA GPU?

For practical local GPU inference, an NVIDIA CUDA-compatible GPU is recommended. MuseTalk’s official documentation reports testing the Windows Gradio interface on an RTX 3050 Ti Laptop GPU with 4 GB VRAM.

4. Which PyTorch version does this installation use?

This installation uses:

torch 2.0.1
torchvision 0.15.2
torchaudio 2.0.2
CUDA 11.8 PyTorch build

These versions match the current official MuseTalk installation instructions.

5. How do I know MuseTalk installed correctly?

Run:

python app.py --use_float16 --ffmpeg_path YOUR_FFMPEG_BIN_PATH

If you eventually see:

Running on local URL: http://127.0.0.1:7860

open that address in your browser.

6. How do I stop MuseTalk?

Return to the Command Prompt running MuseTalk and press:

Ctrl + C

7. Do I need to reinstall everything when I restart Windows?

No. Activate the existing Conda environment:

conda activate MuseTalk

enter the MuseTalk directory, and run app.py again.

Conclusion

Windows installation takes longer than simply cloning the repository because MuseTalk depends on CUDA-enabled PyTorch, OpenMMLab packages, FFmpeg, and several gigabytes of model weights.

The most important things to check are:

1. nvidia-smi works.
2. Python 3.10 is running inside the MuseTalk environment.
3. PyTorch reports CUDA available: True.
4. FFmpeg responds to ffmpeg -version.
5. All required files exist under models.
6. MMPose 1.1.0 imports correctly.
7. huggingface_hub remains below version 1.0.
8. app.py reaches http://127.0.0.1:7860.

Once those checks pass, the MuseTalk Gradio interface should be available locally on Windows.