Lip sync and avatars

LatentSync

Name: LatentSync
Author: ByteDance

Open-weights end-to-end lipsync method built on audio-conditioned latent diffusion, generating mouth movement directly from audio without an intermediate motion representation.

A lip sync and avatar tool by ByteDance · read 2026-08-14 · sourcing high

Resolution: 512x512
Weights: Downloadable
API: No
Commercial use: Permitted subject to conditions

Version 1.6 works at 512x512 and addresses the blurriness of version 1.5, which ran at 256x256. Checkpoints are published on Hugging Face and the code is Apache 2.0.

Commercial use — Permitted subject to conditions

Apache 2.0, which permits commercial use.

Requires consent from the person depicted. The repository notes that its photorealistic demo videos are filmed by contracted models, indicating consent awareness, but the licence itself imposes no likeness-consent obligation, so consent compliance is entirely the deployer's responsibility

The conditions above were extracted from the licence notes recorded during research, and each one shows the sentence it came from so you can check the reading. Read the licence before you rely on any of this. Nothing on this page is legal advice, and this site does not tell you a model is cleared for commercial use.

Specification

Vendor figures, read 2026-08-14Read from ByteDance's own documentation on 2026-08-14. Vendors change these between model versions without announcing it, so check theirs before you commit a shot to a number. Sourcing for this entry is strong: the figures came from pages we retrieved and read. Check the vendor's current documentation

What it takes in: video, audio
What it makes: video
Maximum duration: not established
Maximum resolution: 512x512
Aspect ratios: not established
Controls: lipsync
Weights: Downloadable
API: No
Pricing: Free and self-hosted. Inference needs roughly 8 GB VRAM for v1.5 or 18 GB for v1.6; training needs 23-55 GB depending on resolution and configuration.
Released or updated: 2025-06

Maximum duration and supported aspect ratios are not established here. That means no page we could retrieve stated them, not that the model lacks them.

Strengths

Apache 2.0 with published weights, so it can be self-hosted with no per-second cost
512x512 at v1.6 is competitive with commercial lipsync base resolutions
No vendor lock-in or rate limits

Weaknesses

18 GB VRAM for v1.6 inference puts it beyond consumer GPUs at the low end
No hosted API, SLA or support
No built-in consent or safety controls
Duration behaviour undocumented; training and preprocessing work in 5-10 second segments

Best for

Self-hosted or on-premise dubbing pipelines
Research and fine-tuning
Cost-sensitive high-volume lipsync

Where the sources disagree

Recorded rather than resolved. Picking the tidier number would hide the fact that the vendor's own pages do not agree.

Inference duration is not explicitly capped in the documentation; the 5-10 second segmentation applies to training and preprocessing, so maxDurationSeconds is null rather than 10.

What to understand before using it

Explainers naming the mechanisms this model's own documentation mentions.

Latent space: working at proxy scale names “latent diffusion”
LoRAs and fine-tuning: teaching a model your subject names “fine-tuning”
Diffusion models: generation by denoising names “diffusion”
Training data and copyright: what is not settled names “training”

MCP servers that front ByteDance

Servers in the media MCP directory whose own description or homepage names this vendor. Absence of a server here is not evidence that none exists: most publish no tool list at all.

370.ai, AI Gateway: Video (Seedance 2.0, Wan, HappyHorse), Image, Speech + 100+ Chat Models Video generation
One key, 100+ models, chat with any LLM and generate video, images, speech. Free trial at 370.ai.
Adclip Image generation
Ad creative generation via MCP, copy and static images from a JSON brief. Keyless.
Aetherwave Music
Music, image, video and audio generation across top AI providers - one key, one credit pool.
ArkRoute, Visual AI API Gateway Image generation
Generate images and videos using 28+ visual AI models from OpenAI (GPT Image 2, DALL-E 3), Google (Imagen 4, Veo 3.1), ByteDance (Seedream, Seedance 2.0), and Kuaishou (K
Avocado AI Image generation
Generate AI images, videos, music, SFX & speech in any AI assistant. Results appear inline in chat.
BananaBanana Image & Video Generation Video generation
Generate images & video (Nano Banana, Veo, Omni) pay-as-you-go. No subscription, crypto payments.

2 further servers name ByteDance with lower confidence and are not listed, because a server naming a vendor may be using it for text rather than media.

See it from the vendor

Shotcraft hosts no sample output. Judging a model means watching its own showreel with your own eyes, and a clip re-hosted here would be a clip we chose.

ByteDance's own site and examples

Nearby in the directory

Flawless DeepEditor and TrueSync

Neural performance editing tools for film and episodic post. DeepEditor applies new dialogue to existing shots and refines performances, covering in-v

Lip sync and avatars

Hedra character / avatar platform

Character and avatar video platform that has grown into a unified inference API putting multiple leading image, video and audio models behind one API

Lip sync and avatars

HeyGen Avatar IV (and Avatar V)

Avatar IV is HeyGen's current default avatar engine, driving a photo, consented digital twin or studio avatar from either a typed script (text to spee

Lip sync and avatars

Sources

Read on 2026-08-14. A source that has changed since then will not be reflected here until the next pass.

github.com/bytedance/LatentSync