Best AI Lip Sync Tools (2026): One Voice Line, Four Models

All guides · Best AI avatar video generators

Quick answer: the best AI lip sync tool for making someone in an existing video say new words is LatentSync in Fuser, one of 162 video models on the canvas. It rewrites the mouth and keeps everything else in the take. To make a still portrait talk, use the HeyGen Avatar IV node, which animates the head, face and hands to your audio. If you're generating the shot anyway, MiniMax H3 takes your voice file as a reference and builds the scene around it, speech included. Seedance 2.5 also takes reference audio, but it re-voices the line rather than using your recording, and it refused our photoreal face. To transfer a performance from a video of an actor instead of an audio file, use the Runway Act-Two node.

The test: one face, one voice line

All four models got the same inputs. The face is the potter from our Veo 3.1 tests: the first six seconds of that clip with its sound removed, plus one frame from it at 3.5 seconds for the models that start from an image. The voice line came from ElevenLabs Eleven v3 (voice "Laura"): "Keep your palms wet, and let the wheel do the work. Big breath... almost there." It runs 6.3 seconds and has three pauses, and the pauses are the useful part. A good lip sync closes or relaxes the mouth when the voice stops.

To check what each model did with the sound, we compared the audio it returned against our file. HeyGen's matched exactly, LatentSync's was the same apart from encoding, and H3's was our line with workshop sound mixed underneath. Seedance's did not match at all: the words and pauses were the same, but it was a new recording.

LatentSync and HeyGen: the dedicated tools

LatentSync is ByteDance's open lip-sync model, an audio-conditioned latent diffusion model that uses Whisper to read the speech (GitHub). You give it a video and an audio file. It left the rest of the take alone: the potter still looks down at the wheel, shapes the vase and glances up, just as in the source. Only the mouth changed. Where the original Veo take had her say "Almost there" at about three seconds, her mouth now rests through our pause. The sync is easiest to read when she faces the camera. When her head tips down toward the wheel, the mouth shapes get smaller.

Our first run came back at 5.76 seconds, shorter than both the 6-second video and the 6.3-second line, so the last word was cut. We padded the audio with silence to 7 seconds and ran it again, and the second run returned 6.4 seconds with the whole line. The clip above is that second run. When the audio is longer than the video, the node's Loop Mode (pingpong by default, or loop) decides how the video is extended. The node also has a Guidance Scale from 1 to 2 and a Seed. Even counting the rerun, the two runs together cost a small fraction of one Seedance 2.5 clip.

HeyGen Avatar IV starts from one image. HeyGen describes an "audio-to-expression engine" that adds "head tilts, natural pauses, subtle cadences, and micro-expressions" (HeyGen). In our run it did all of that: the head moves, she blinks, the hands shift on the clay, and she smiles through the pauses. The face stayed close to the photo. It returned 1920 × 1080 at 25 fps, the same length as the audio. This is on the Stable talking style, which is the node's default. Expressive animates more.

Pick by your starting point. If the shot, the camera move and the acting already exist and only the words need to change, as in a dub, a fixed line or a translated ad, use LatentSync. If all you have is a still, use HeyGen. Per second of output HeyGen costs about twenty times what LatentSync does.

Try it on your clip: LatentSync for a take you already have, or HeyGen Avatar IV for a single photo.

HeyGen lip sync in Fuser

HeyGen Avatar IV lip syncs one portrait to your own audio. Wire a recording, or a line from an ElevenLabs node, into the node's Audio input, and that file drives the performance: it sets the clip length and overrides the script and voice settings. In our audio test earlier on this page, HeyGen returned our ElevenLabs line unchanged at 1920 × 1080 and animated the potter to it, smiling through the pauses.

If you don't have a recording, leave Audio empty, type the line into Prompt and pick one of the node's 100-plus HeyGen voices. HeyGen then generates the speech and the lip sync in one run. The clip here is that mode: the same potter frame, the script "Trim the foot when it's leather-hard. Too soon, it tears. Too late, it chips." and the Calm Chloe voice. It came back as a 5.2-second 1080p clip. A Whisper transcript of its audio matched the script word for word, and the face stayed close to the frame while her head and clay-covered hands moved with the speech.

Resolution runs from 360p to 1080p, with 720p as the default, so you can check a script at 360p and render the final at 1080p. Aspect ratio can be 16:9, 9:16, 4:5, 5:4, 1:1 or auto. Talking Style is Stable (the default, restrained) or Expressive (bigger brows and gestures).

When to use LatentSync instead. HeyGen builds the whole performance from a still, so the head movement, the hands and the expression are its own. If you already have the take and only the words should change, LatentSync keeps the camera move, the acting and the timing, and redraws only the mouth, at about a twentieth of HeyGen's cost per second.

MiniMax H3 and Seedance 2.5: audio as a reference

These are video generators, not lip-sync tools, but both accept reference audio next to reference images. We gave both the same prompt: the woman from the image "looks into the camera and says the words in Audio 1, her lips matching Audio 1 exactly. Use Audio 1 as her voice", with a static medium shot.

MiniMax H3 used our recording as the dialogue track. Its own prompt expansion said so, and the audio it returned is our line in the same place, with wheel hum added underneath. The difference from HeyGen is that H3 regenerates the whole shot: her hands work the vase, she looks at the camera, and she smiles at the end. The face drifted a little from the reference frame, which is the price of generating new video. MiniMax's own example of this feature is "have the character in Image 2 sing, with the vocals matching Audio 3" (MiniMax). In Fuser, reference audio clips must be 2 to 15 seconds each and 15 seconds combined, cited as Audio 1, Audio 2 and so on, and audio can't be the only reference. Our 768p run returned 1344 × 768.

Seedance 2.5 rejected our first request. The model returned a content-policy error on the image: it "may contain likenesses of real people". That happened even though the potter is AI-generated. So we made a clay-animation version of the same frame with GPT Image and ran the same prompt again. This time Seedance delivered the line with the same words and close to our pauses, and the mouth follows the speech. But the audio was a new performance, not our file. The reference audio guided the rhythm, timing and voice, but the output isn't a copy of it. Budget for it, too. Seedance 2.5 bills per token, and our 6-second 720p clip cost many times more than the dedicated tools.

Use H3 when the scene doesn't exist yet and you want the character to speak your exact recording in it. Use Seedance's audio reference for stylised characters, where hitting the rhythm matters more than keeping the exact recording.

Try it on your clip: MiniMax H3 to build a new shot around your recording, or Seedance 2.5 for a stylised character.

Runway Act-Two: performance, not audio

Runway Act-Two doesn't take an audio file. It takes a video of a person performing, 3 to 30 seconds long, and moves their facial expression, and optionally their gestures, onto a character image or video (Runway). If you can film yourself saying the line, you get your delivery, not a model's guess at it. See the Act-Two guide for how to shoot the take.

Try it on your clip: Runway Act-Two with a 3 to 30 second video of someone performing the line.

How to lip sync a video with AI in Fuser

This is the lip sync AI workflow we used for the LatentSync clip above, from a finished take to a new line.

1. Add the video. Drop your clip onto the canvas to make a Video node. Choose a take where the face is visible and roughly toward the camera when the key words land.

2. Add the voice. Drop a recording onto the canvas as an Audio node, or generate the line with an ElevenLabs TTS node. Keep it to dialogue only, and add music after the lip sync.

3. Wire both into LatentSync. Connect the video to the LatentSync node's Video input and the voice to its Audio input. Guidance Scale runs from 1 to 2. Loop Mode (pingpong or loop) only matters when the audio is longer than the video.

4. Pad the audio if the end is cut. Compare the output's length with the voice line. If the last word is missing, add half a second or more of silence to the end of the audio and run it again. That fixed our first run, which came back at 5.76 seconds.

5. Export. Download the synced clip, or keep wiring: captions, a transcript check, or an upscaler, as below.

Tips that held up in our runs

• Pad the audio. LatentSync's first output came back shorter than both inputs and cut the last word. Half a second or more of trailing silence fixed it.

• Face the camera for the key words. LatentSync's mouth shapes were clearest when the potter looked up. Pick takes where the face is visible when the important words land.

• Leave pauses in the script. Pauses show sync problems quickly, and they give HeyGen and H3 room to act.

• Keep the voice file clean. LatentSync reads the speech through Whisper audio embeddings, so give it dialogue only and add music after the lip sync.

Build it on one canvas

In Fuser the voice, the face and the lip-sync models sit on one canvas. Generate or record the line, wire it into LatentSync, HeyGen and H3 side by side as in the image at the top, and keep the version that suits the shot. From there you can add captions with VEED subtitles, check the words with Whisper, or send the clip to an upscaler (best AI upscalers). To run the same steps on the next line, save the graph as a Recipe. For presenter-style videos made from scratch, see best AI avatar video generators. For video models that generate their own dialogue from a prompt, see best AI video generators with audio.

HeyGen and LatentSync node settings and the HeyGen script test last verified October 23, 2026. The other tests ran on 28 September 2026.