Code under Apache 2.0 and weights under the OpenRAIL++ license, which allows commercial use within its use restrictions. Self-hosting needs a GPU with at least 18GB of memory for version 1.6, against 8GB for version 1.5, and the repository ships the weights, a training recipe and an inference script. In your browser it runs on cnaps studio, included in the basic plan.
It changes one thing in a video you already have: the mouth. Give it a talking clip and a new voice track, and it regenerates only the lips to match the new audio, leaving the face, the background and the camera move as they were. That makes it a dubbing and re-voicing tool rather than a video generator.
The cost is a little softness. The new mouth is drawn at 512 by 512 pixels on the face, so it can look slightly softer than the rest of a sharp close-up, and every frame needs one visible face. Eight video generation models sit in the cnaps studio catalog, and this is the one for re-syncing footage that already exists.
What it actually is
An audio-conditioned latent diffusion model from ByteDance. A small speech encoder turns the voice into embeddings that steer an image diffusion network through cross-attention, and the network repaints a masked mouth region using a reference frame of the same face. During training, a lip-sync judge network the authors call StableSyncNet scores every result, which is where most of the sync accuracy comes from.
Version 1.6, released in June 2025, was retrained on 512 by 512 face crops to reduce the blur of earlier versions. It works in windows of 16 frames and keeps the source video's frame rate.
In cnaps studio you connect a video and a sound file. The voice can run up to 60 seconds and is muxed into the result. Steps run from 5 to 50 per window, default 20, and an advanced guidance setting runs from 1.0 to 3.0, default 1.5, where higher values exaggerate articulation. A 13-second clip takes about a minute and a half at the default.
Try LatentSync 1.6 in cnaps studio →
The published numbers
The paper scores the original LatentSync, trained at 256 by 256, against four other open lip-sync methods on two talking-head test sets. FID and FVD measure how real frames and motion look, SSIM how closely frames match the original, LMD how far mouth landmarks drift, and Sync how well lips match the audio. Lower is better for FID, LMD and FVD; higher for SSIM and Sync.
Those scores were the best of the five methods in all ten columns. The authors also report that StableSyncNet lifted the judge network's own accuracy on HDTF from 91% to 94%.
Read them as a strong starting point measured by the authors, on an older version at half the face resolution. No scores have been published for 1.6 itself.
How it compares to the other talking-video models here
These four share no published test, so the table maps what goes in. The first question is whether you start from footage, a still, or nothing at all.
Reach for LatentSync 1.6 when the video exists and only the words change. Reach for LTX-2.5 Audio-to-Video to animate a portrait to a voice, MiniMax H3 to build a scene from references, and LTX-2.5 when the model can invent the speech too.
What to chain it with
A video and a voice in, the same video with matching lips out, so it usually finishes a chain.
The natural chain is dubbing. Speech Recognition (Clips) transcribes the original, Gemma 4 31B translates the script, Qwen3-TTS or Qwen3 Voice Design speaks it in the new language, and LatentSync 1.6 fits the lips. For a speaker who does not exist yet, LTX-2.5 can generate the footage first, as in the example above.
Open LatentSync 1.6 in cnaps studio, sync one short close-up at the default 20 steps, and check fast syllables and teeth before running a whole video.
FAQ
What does LatentSync 1.6 need to run locally?
A GPU with at least 18GB of memory for version 1.6, against 8GB for version 1.5. The repository provides the weights and an inference script, and the weights use OpenRAIL++, which allows commercial use within its restrictions. In cnaps studio it runs in the browser, included in the basic plan.
Why did my run fail?
Every frame needs one visible face. The studio aligns the face in each frame before generating, and a frame with no face, such as a cutaway or an empty shot, stops the run. Trim those frames first, or split the video and sync only the parts where the speaker is on screen.
What happens if the voice is longer than the video?
The output follows the sound, up to 60 seconds. When the video is shorter, the studio plays it forwards and then backwards to cover the gap, which suits a steady talking shot but looks odd with visible motion. For long narration, supply a video at least as long as the voice.
