What is SadTalker and is it still good in 2026?

SadTalker has 14k stars and 2.7k forks on GitHub, which makes it one of the most-starred open-source talking head models ever released. SadTalker is a 2023 deep learning model that takes a single portrait photo and an audio clip and generates a talking head video, with the head motion and facial expressions driven by the audio. It was published at CVPR 2023 and became the default free way to make a still image speak. The same job runs hosted on sync-3, which turns one photo into a talking video with image-to-video and dubs it across 95+ languages.

If you are researching SadTalker today, you are usually deciding one of two things: whether to run the open-source model yourself, or whether something newer does the job better. This page explains what SadTalker is, how it works, where it shows its age in 2026, and what to use when you outgrow it.

SadTalker turns one photo and an audio clip into a talking head video

SadTalker generates a talking head video from a single still image of a face plus any audio track. You give it a photo and a voice recording, and it produces a video of that face speaking the audio, with lip motion, blinks, facial expression, and head movement it invents from the sound. The paper behind it, “SadTalker: Learning Realistic 3D Motion Coefficients for Stylized Audio-Driven Single Image Talking Face Animation,” was published at CVPR 2023 by Zhang et al.

Timing made it popular. SadTalker landed in early 2023, right as Stable Diffusion communities were looking for ways to animate the faces they were generating, and it shipped as a Stable Diffusion WebUI extension alongside the standalone code. Generate a portrait in one tab, make it talk in the next. That workflow put it in front of a much larger audience than most research code ever reaches, and it is a big part of why the repo sits at 14k stars.

How SadTalker works: 3D motion coefficients instead of 2D warping

SadTalker’s core idea is to generate the motion of a talking head as 3D face model coefficients rather than warping the image directly in 2D. Earlier single-image methods learned a flat 2D motion field from audio, which tangled head movement, expression, and identity together, so faces drifted, distorted, or stopped looking like the person. SadTalker splits the problem into pieces of an explicit 3D face representation (a 3DMM, or 3D morphable model), generates each piece separately, and only then renders the final video.

Two modules do the generating. ExpNet learns facial expression coefficients from the audio, trained against both the motion coefficients and rendered 3D faces so the lips actually track the words. PoseVAE, a conditional variational autoencoder, synthesizes the head movement, which is why a SadTalker clip nods and shifts naturally instead of holding a frozen skull. The generated coefficients then drive a 3D-aware face renderer that produces the video, at a face resolution of 256 or 512 pixels, with an optional GFPGAN pass to sharpen the result. The code is Apache 2.0 licensed, so you can use it commercially.

SadTalker is not a lip sync model for existing video

SadTalker animates a still photo; it does not re-sync the lips in footage you already have. That is the most common confusion around it. A lip sync model like Wav2Lip takes an existing video of someone talking and edits the mouth to match new audio, keeping everything else about the performance. SadTalker starts from zero: one frame, no motion, and it invents the entire performance from the audio.

The distinction decides which tool you need. If you have a video and want the mouth to match a new voice track (a dub, a translation, a re-record), SadTalker is the wrong tool no matter how good the output looks. If you have only a photo, SadTalker was for years the best free answer, which is exactly the gap image-to-video on sync-3 now covers hosted, and sync-3 handles the existing-video case too.

Where SadTalker shows its age in 2026

SadTalker has not seen significant development since 2023, and the gap shows against current models. As of July 2026, the repo’s own update log stops in mid-2023 and more than 600 issues sit open. A few limits show up quickly in real use:

• Low resolution. The face renders at 256 or 512 pixels and is enhanced back up with GFPGAN, which produces the slightly waxy, over-smoothed look SadTalker clips are known for. There is no native HD or 4K output.
• Head-and-shoulders framing. It animates a cropped face. Shoulders, hands, body language, and anything else in frame stay frozen or need separate tooling.
• Audio in, nothing else. SadTalker syncs to an audio file you already have. No text-to-speech, no translation, no voice cloning, so a dubbing workflow means stitching several tools together.
• The uncanny drift. Invented head motion loops and repeats on longer clips, and expressions read flat next to 2026-era models that generate every frame with full-scene context.
• You run and maintain it. It is a research codebase with Python dependencies and GPU requirements, not a product, and with development stalled you also inherit its unfixed bugs.

None of this makes SadTalker a bad choice for a hobby project or a research baseline. It just means the bar for a talking head video has moved a long way since 2023.

SadTalker vs modern image-to-video models

The gap between SadTalker and a current hosted model is resolution, realism of the generated performance, and what happens around the generation step. Specs below are from each project’s own published sources.

Model	Input	Resolution	Works on existing video	Voice cloning + dubbing	How you run it
SadTalker	One photo + audio	256 or 512px face, enhanced	No, image only	No	Self-hosted open-source code
sync-3	One photo + audio, or any video	Up to 4K at 60fps	Yes, zero-shot on any footage	Yes, one-pass across 95+ languages	Hosted: web, API, MCP, plugins

Sources: SadTalker, sync-3 as of July 2026.

sync-3 generates the talking video from a whole-scene read of the face rather than a cropped 3D head model, which is why the output holds up at modern resolution and why the same model works on a photo or on real footage. Give it one still and image-to-video builds the full performance: mouth, expression, head motion, every frame generated together. Give it a video, live-action, animation, or AI-generated, and it lip syncs zero-shot while preserving how that person speaks. And the pipeline SadTalker leaves you to assemble is built in: translation, voice cloning, and lip sync run in one pass across 95+ languages, with an average processing time under 3 minutes.

How to choose between SadTalker and a hosted model

Start from what you are trying to ship. If you want a free research baseline you fully control, or you like tinkering inside the Stable Diffusion ecosystem, SadTalker still runs. If anyone else is going to watch the output, the resolution and realism gap is the deciding factor.

Your job	Best fit	Why
Experiment with talking heads for free	SadTalker	Open-source, Apache 2.0, runs in the SD WebUI
Turn a photo into a video people will watch	sync-3	Full performance generation, no waxy enhancer look
Lip sync or dub footage you already have	sync-3	SadTalker cannot edit existing video at all
Speak in another language	sync-3	One-pass translation, voice cloning, and lip sync in 95+ languages
Avoid GPUs, dependencies, and stalled repos	sync-3	Hosted in the browser, API, and MCP clients

For a wider look at the open-source options around SadTalker, see the best free and open-source lip sync tools.

Frequently asked questions

Try image-to-video on sync-3 in the browser: upload one photo and an audio clip and get a talking video back, usually in under 3 minutes, with 3 free videos a month to test it on.