LatentSync is ByteDance's Apache-2.0 lip-sync model: it changes mouth motion in a supplied video to match new audio, using Stable Diffusion with Whisper-based audio conditioning. Pick it when you have footage and permission to alter it. Version 1.5 needs 8 GB VRAM, 1.6 needs 18 GB; it decides nothing about consent, identity or publication rights.
Other good fits
Projects that need an Apache-2.0 licensed, inspectable model rather than a black-box cloud service.
+ Pros
Read pros
LatentSync is ByteDance's open-source lip-sync model: given a video and a new audio track, it changes the mouth motion in the video to match the new audio, using a Stable Diffusion-based approach with Whisper audio embeddings for the audio-to-mouth mapping. It's released under Apache-2.0, so you can inspect and self-host it, unlike a closed cloud lip-sync service.
Pick it when…
You have your own footage and audio, and the rights to modify both, for example replacing a presenter's dubbed line.
You want a self-hosted, inspectable model instead of sending footage to a third-party cloud service.
Your hardware meets the VRAM requirement for the version you need (8 GB for 1.5, 18 GB for 1.6).
− Cons
Read cons
LatentSync changes mouth motion in footage you supply; it says nothing about whether you have consent to alter that footage, whether the identity shown is being represented accurately, or whether the final output needs a disclosure label before publication. The version you use, the source clip's quality, and the inference settings all materially affect visible artifacts and sync accuracy, so a good result on one clip doesn't guarantee one on another.
Pick something else when…
You need to generate new video from a prompt, not resync existing footage.LTX-Video or HunyuanVideo generate scenes rather than editing supplied video.
You'd rather use a managed cloud workflow than self-host.Kling covers a proprietary cloud video path with its own trade-offs.
You need this for a live or interactive conversation, not a fixed clip. Look at a realtime video-agent tool instead; LatentSync works on supplied files, not live streams.
Avoid if
You don't have 8-18 GB of VRAM or the consent and rights to alter the footage in the first place.
Context
The project has two current versions: 1.5 (March 2025) and 1.6 (June 2025, trained on higher-resolution 512x512 video for less blur). Use the version whose documented requirements and quality trade-offs match your project; don't assume settings tuned for one version carry over cleanly to the other.
Requirements
Run it through the included Gradio web interface (python gradio_app.py) or the command-line ./inference.sh script. Inference needs 8 GB VRAM minimum for 1.5 or 18 GB for 1.6; training (if you go that far) needs 20-55 GB depending on resolution and stage. Inference parameters like steps (20-50) and guidance scale (1.0-3.0) trade speed against lip-sync accuracy and visual quality.
Licence & cost
The code and weights are free under Apache-2.0. You pay for GPU time: renting cloud GPU hours with the required VRAM, or buying/using hardware that already meets it, plus storage for footage and checkpoints. This review doesn't make a legal determination about your specific footage, voices or output; that's a rights question you settle before running anything, not something the tool settles for you.
First task
Documentation-based suggestion — not run here.
Plan a consent-checked dub before touching any real footage.
Pick a short clip you have explicit permission to modify, with a presenter who has agreed to the dub.
Write down exactly what the replacement audio line will say.
Record or generate that replacement audio.
Run LatentSync on the clip and audio, then have a second person check lip alignment and confirm the result doesn't misrepresent what was actually said.
Done when the synced clip is reviewed by someone other than you for alignment quality and for whether it needs a disclosure label before anyone sees it.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on ByteDance's LatentSync GitHub repository, the 1.6 model card and the technical report, checked 23 September 2026. No footage was processed and no inference was run for this review.
Align supplied audio with mouth motion in existing face video.
Replace dialogue audio for permitted video projects.
Review generated motion before sharing or publishing.