SadTalker vs MuseTalk — same-input portrait comparison
Recorded: 28 October 2026
Page: https://sadtalker.ai/sadtalker-vs-musetalk

Hardware: NVIDIA GeForce RTX 3090, 24,576 MiB VRAM.
Input: shared synthetic portrait, 512 x 474 PNG; shared 8.000-second mono
16 kHz PCM WAV. Download portrait.png and speech.wav alongside this file.
The files were prepared once, then passed unchanged to both models.
Input and output SHA-256 hashes, exact revisions and package versions are
in results.json. Each model has its own isolated Python environment.

SETUP
Follow each official project's setup and model-download instructions:
https://github.com/OpenTalker/SadTalker
https://github.com/TMElyralab/MuseTalk

The recorded environments use:
SadTalker: Python 3.10, PyTorch 2.1.2 + CUDA 12.1.
MuseTalk: Python 3.10, PyTorch 2.0.1 + CUDA 11.8, MMCV 2.0.1.
See results.json for the complete version strings and upstream revisions.
FFmpeg must be available. The recorded environment was Linux x86_64.

SADTALKER
From the SadTalker repository, with its environment active, place the shared
files in an inputs directory and run:

python inference.py --source_image inputs/portrait.png --driven_audio inputs/speech.wav --result_dir results/comparison --size 256 --preprocess full --still --batch_size 2 --expression_scale 1 --pose_style 0

Enhancer and background enhancer are omitted (off).

MUSETALK 1.5
From the MuseTalk repository, with its environment active, place the same
shared files in an inputs directory. Save comparison.json containing:

{
  "task_0": {
    "video_path": "inputs/portrait.png",
    "audio_path": "inputs/speech.wav",
    "result_name": "result.mp4"
  }
}

Run:

python -m scripts.inference --inference_config comparison.json --result_dir results/comparison --unet_model_path models/musetalkV15/unet.pth --unet_config models/musetalkV15/musetalk.json --whisper_dir models/whisper --version v15 --fps 25 --batch_size 2 --extra_margin 10 --use_float16 --ffmpeg_path /usr/bin

MEASUREMENT
Two fresh-process renders per model ran sequentially after setup, in this
order: SadTalker 1, MuseTalk 1, SadTalker 2, MuseTalk 2. All model files were
local. Wall-clock timing used Python time.monotonic() around subprocess.run().
It includes loading, face preparation, generation, composition and MP4 export.
Installation, downloads, input preparation, validation and browser-copy encoding
are outside the timed interval. Fresh process means models are loaded again;
operating-system disk caches may remain warm between runs.

A separate thread sampled this command every 0.5 seconds:
nvidia-smi --query-gpu=memory.used,utilization.gpu --format=csv,noheader,nounits

The reported peak is whole-device sampled occupancy, including baseline use.
It describes this run configuration; use it as an observation when planning
memory, rather than a minimum GPU-memory requirement.

VALIDATION
ffprobe confirmed video and audio, 512 x 474 dimensions, 25 FPS, 200 frames
and 8.000-second video/audio duration for each model's exports.
Every exported MP4 completed this decode check:
ffmpeg -v error -xerror -i result.mp4 -f null -

PLAYBACK COPIES
The embedded players use H.264 browser copies made after measurement:
ffmpeg -i native.mp4 -c:v libx264 -crf 18 -pix_fmt yuv420p -c:a copy -movflags +faststart browser.mp4

No face enhancement was added during this conversion. Native exports are
available as sadtalker-native.mp4 and musetalk-native.mp4.
Run 1 supplies each embedded example. speech.vtt provides English captions.

SCREENSHOTS
The SadTalker screenshot comes from our community still-mode workflow guide:
https://sadtalker.ai/sadtalker-still-mode
The MuseTalk screenshots come from our community ComfyUI guide:
https://musetalkai.com/comfyui/
They explain the interfaces and are separate from the CLI measurement above.