AI Lip Sync Tools: How They Work and What They Cost (2026)

AI lip sync tools take a face, either a still photo or existing video, plus an audio track, and generate new frames where the mouth moves in time with the words. That one description covers phone apps that make a selfie sing, dubbing services that re-voice an interview in another language, and developer APIs that turn a single portrait of a presenter into a narrated clip. From the outside they look alike, but they differ a lot in what they accept, how much control you get and how they charge. Picking the wrong kind is the most common reason people give up after one bad result.

This guide explains how the technology works, sorts the options into groups you can actually choose between, is honest about what free tiers give you, and walks through a working script that produces a lip sync video from one image and one audio file, with the cost worked out to the credit.

What are AI lip sync tools?

At their core, AI lip sync tools solve an alignment problem. Speech is a sequence of sounds, and a face makes a matching sequence of mouth shapes. A tool that gets the alignment right produces a clip nobody questions. One that gets it wrong feels off within a second, because people read faces constantly without thinking about it, and a mouth that closes a frame late is enough to break the illusion.

Before machine learning, matching mouths to dialogue meant an animator keying mouth shapes by hand or rigging a 3D face. Modern lip sync video AI learns the mapping from large amounts of footage of people talking, so your job shrinks to supplying three things:

• A face source. A single photo (a portrait, an illustration, a stylised character) or a video of someone already talking.
• An audio source. A recorded voice, a file from a text-to-speech (TTS) engine, or a song.
• Optional guidance. A text prompt describing the scene or mood, a mask image that marks a region of the frame, and output settings such as resolution.

The result is a new video file. Some tools only rewrite the mouth area and paste it back onto your original footage. Others generate the whole frame, adding head movement, blinks and sometimes gestures, which is why a still photo can come out looking like a real recording.

How does lip sync video AI work?

Every system in this category follows roughly the same pipeline, even when the internals differ.

1. The audio becomes a timeline of sounds

The model first breaks the audio into short slices and extracts features from each one. Two terms come up here. A phoneme is the smallest unit of sound in speech, like the “p” in “pat”. A viseme is the visible mouth shape that goes with a sound. Several phonemes share a viseme: “p”, “b” and “m” all close the lips, which is why lip reading is hard and why lip sync models can get away with a smaller vocabulary of shapes than the full set of sounds.

2. The face is located and described

Next the system finds the face in the image or each video frame, tracks landmarks such as the corners of the mouth and the jawline, and builds an internal description of identity: skin, teeth, lighting, angle. Keeping identity stable while the mouth changes is the hard part. Weak models produce a blurry lower face or teeth that shift between frames.

3. New frames are generated to match the audio

Most current tools use diffusion models, a class of generative model that starts from noise and removes it step by step until an image appears, guided by conditions such as the audio features and the reference face. The model produces frames at a set frame rate, the number of images shown per second of video. At a typical rate, every second of speech needs dozens of consistent frames, which is why lip sync is priced per second and why long clips take longer to render.

Mouth-only editing vs whole-frame generation

Two families of AI lip sync tools fall out of this pipeline. Mouth-only editors keep your original video and regenerate just the lower face. They are the natural fit for dubbing and fixing a flubbed line, because the rest of the performance stays real. Whole-frame generators create the entire video from a still image and the audio. They can make a photo talk, sing or present, and they add natural head and body motion, but you hand over more creative control to the model.

Types of AI lip sync tools compared

Search results for this topic mix very different products into one list. Sorting them by input and use case makes the choice much easier.

Type	What you give it	Best for	Main trade-off
Mobile and web photo apps	A selfie or photo plus a song or recording	Social clips, memes, quick tests	Watermarks, short limits, little control
Avatar platforms	A written script, a stock or custom avatar	Training videos, sales outreach, explainers	Polished but templated look, tied to one editor
Video re-sync and dubbing services	Existing footage plus new audio	Translating videos, replacing lines in post	Needs good source footage of the face
Open-source local tools	Your own files, your own GPU	Privacy, tinkering, zero per-clip cost	Setup time, hardware, uneven quality
Model APIs	Image and audio URLs over HTTP	Apps, batch production, automation	You write a little code

Photo apps

Apps in the phone stores and web tools from creative suites such as OpenArt, Dzine and Magic Hour make a still image sing or talk in a few taps. They are the fastest way to see what the technology can do, and for a one-off social post they are often enough.

Avatar platforms

HeyGen is the name many people search for when they want HeyGen AI lip sync specifically. It builds avatar videos from a script and offers video translation with lip sync. D-ID and Akool sit in the same broad space. These suit teams who want a finished editor with avatars, voices and templates in one place, and who are happy to work inside that product.

Re-sync and dubbing services

Tools such as Sync focus on taking real footage and making the mouth match a different audio track, often in another language. If you already have good video of a speaker and need it in Spanish, this is the category to look at, not a photo animator.

Open-source and local

Open-source models such as Wav2Lip, plus community-built desktop interfaces that wrap them, let you lip sync entirely on your own machine. You pay nothing per clip, your files never leave your computer, and you can modify the code. You also handle installs, drivers and a GPU, and older open models tend to blur the mouth at higher resolutions.

Model APIs

If you are building a product, generating clips in bulk or wiring video into an agent, you want an API: send an image and an audio URL, get a video back. This is where Unicorn API fits. One key gives you lip sync alongside image-to-video and keyframe models, billed in the same credits, so the talking segments and the cutaway shots in a video come from the same account.

Are free AI lip sync video generators any good?

“AI lip sync video generator free” is one of the most common searches around this topic, and the honest answer is: good enough to try, rarely good enough to ship.

Free tiers of hosted AI lip sync tools usually come with one or more of these limits:

• Watermarks burned into the video.
• Length caps that stop you well short of a full explainer or song.
• Lower resolution, which shows up first in teeth and lips.
• Queues that put free jobs behind paying ones at busy times.
• Usage terms that may let the service keep or reuse your uploads. Read them before you upload a client’s face.

Open-source tools are the genuinely free route. They are a sensible choice if you have a capable GPU, care about privacy, and enjoy tuning settings. Budget an afternoon for setup, and expect to spend time on face cropping and audio cleanup to get clean results.

Free is fine for testing whether a photo works at all, making memes, or learning what kind of audio gives good mouth motion. Once a video carries your brand or a client’s, the watermark, the resolution ceiling and the terms usually push you to a paid tool. The useful question then becomes cost per finished second, which is easy to answer with per-second pricing.

How to make a lip sync video with an API

On Unicorn API the model built for this job is OmniHuman 1.5, a ByteDance video model released on 2026-06-15. Its inputs are a still image, an audio file, an optional prompt and a few optional settings, which is exactly the shape of the photo-to-talking-video problem. Here is the full process.

Step 1: Prepare the image

Use a front-facing or slightly angled portrait with the whole face visible and the mouth unobstructed. Even, soft light works better than harsh shadows across the jaw. The face should take up a decent share of the frame: a tiny face in a wide shot leaves the model very few pixels to animate.

Step 2: Prepare the audio

Clean speech with no music bed gives the sharpest mouth motion. If you only have text, generate a voice file with a TTS model first (browse the full model catalog for audio models), then use that file. For songs, isolate the lead vocal and use that stem. Trim silence at both ends, because the clip is billed per second.

Step 3: Host both files

The model takes image_url and audio_url, so both files need to be reachable over HTTPS. Any object storage bucket with public or signed URLs works.

Step 4: Create the job and poll for the result

Create an API key at your keys page and set it as UNICORN_API_KEY. Then run this Python script:

import os
import time
import requests

API = "https://api.unicornapi.net"
HEADERS = {"Authorization": f"Bearer {os.environ['UNICORN_API_KEY']}"}

payload = {
    "model": "omnihuman-1-5",
    "input": {
        "prompt": "A presenter in a bright home office speaks warmly to the camera, natural head movement",
        "image_url": "https://example.com/presenter.png",
        "audio_url": "https://example.com/voiceover.mp3",
    },
}

# 1. Create the job
resp = requests.post(f"{API}/v1/jobs", json=payload, headers=HEADERS, timeout=30)
resp.raise_for_status()
job_id = resp.json()["id"]
print("Job created:", job_id)

# 2. Poll until it finishes
while True:
    r = requests.get(f"{API}/v1/jobs/{job_id}", headers=HEADERS, timeout=30)
    r.raise_for_status()
    job = r.json()
    status = job["status"]
    if status == "succeeded":
        print("Video ready:", job["outputs"][0]["url"])
        break
    if status in ("failed", "canceled"):
        raise RuntimeError(f"Job ended with status {status}: {job}")
    time.sleep(5)

The same request with curl, if you want to test from a terminal first:

curl -X POST https://api.unicornapi.net/v1/jobs \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "omnihuman-1-5",
    "input": {
      "prompt": "A presenter in a bright home office speaks warmly to the camera",
      "image_url": "https://example.com/presenter.png",
      "audio_url": "https://example.com/voiceover.mp3"
    }
  }'

# Then poll with the returned id:
curl https://api.unicornapi.net/v1/jobs/JOB_ID \
  -H "Authorization: Bearer $UNICORN_API_KEY"

OmniHuman 1.5 also accepts seed, mask_url, pe_fast_mode and output_resolution. A fixed seed helps you reproduce a result you liked while you tweak the prompt. The exact accepted values for the other fields are in the API documentation. If you would rather click than code, the same models run in the video playground, and if you work in Cursor or Claude Code you can call them through the MCP integration.

What do AI lip sync tools cost per second?

OmniHuman 1.5 costs 27 credits per second of output video. There is no duration field in its input, so the clip length follows what you send, which in practice means your audio. Monthly plans give credits at $0.01 each (Starter $25, Creator $50, Studio $100, Pro $200), and top-ups cost $0.015 per credit and never expire. See plans and credit prices for details.

Clip length	Credits	At plan rate ($0.01)	At top-up rate ($0.015)
10 seconds	270	$2.70	$4.05
30 seconds	810	$8.10	$12.15
60 seconds	1,620	$16.20	$24.30

Put another way, the Starter plan’s 2,500 credits cover about 92 seconds of lip sync, and the Studio plan’s 10,000 credits cover about six minutes. Before a big batch, call POST /v1/models/price to get the exact price of a request before running it.

Worked example: a 45-second product explainer

A typical short explainer alternates between a presenter talking and shots of the product. Say you want three 10-second talking segments and three 5-second cutaways:

• Talking segments: 3 × 10 s with OmniHuman 1.5 = 810 credits.
• Cutaways: 3 × 5 s with Grok Imagine Video 1.5 at 720p. Each costs 22.5 credits, rounded up to 23, so 69 credits.
• Total: 879 credits, about $8.79 at the plan rate.

Grok Imagine Video 1.5 does not take an audio track, so it is not a lip sync model. It animates a start frame for 1 to 15 seconds at very low cost per second, which makes it ideal for B-roll. At 480p a 5-second cutaway is just 12 credits. It can be busy at peak times, so build in a retry. If you need longer single shots, Grok Imagine Video 1.0 runs 6 to 30 seconds at the same per-second rate.

Two other models are worth knowing for the non-talking parts. Kling 4.0 Keyframe To Video accepts first-frame, last-frame and keyframe images, at 28.5 credits per second at 720p or 42.5 at 1080p, so a 5-second 720p transition comes to 143 credits. That is useful for a controlled move from a product shot to your presenter’s opening frame. PixVerse V6 Max Image To Video runs from 4.8 credits per second at 360p up to 22.4 at 1080p with its audio switch on. Our PixVerse V6 Max price and API guide covers it in detail.

Common mistakes with AI lip sync tools

Most bad results come from the inputs, not the model. These are the problems that show up again and again:

1. Music under the voice. Background tracks and harmonies blur the timing signal. Lip sync the dry voice, then add music in your editor.
2. Faces in profile. A side-on face hides half the mouth. Stay within a gentle three-quarter angle.
3. Something covering the mouth. Microphones, hands, hair and heavy shadows all confuse the mouth region. Crop or reshoot.
4. A tiny face. If the face is a small part of a wide frame, crop tighter before upload.
5. One giant generation. A two-minute clip that goes wrong at 1:40 means paying for the whole thing again. Split audio at sentence boundaries and generate segments you can replace one at a time.
6. A prompt that fights the audio. A calm, slow voiceover paired with a prompt asking for energetic gestures produces odd body language. Match the prompt to the delivery.
7. Skipping the price check. Per-second billing is predictable, but only if you check length first. Trim leading and trailing silence.
8. Using a likeness without consent. Animating a real person saying something they never said can break platform rules and, in some places, the law. Get written permission and label synthetic speech where viewers could be misled.

How to choose between AI lip sync tools

The right pick follows from your input and your output, not from a ranked list:

• You have a photo and want a quick social clip: a free photo app is fine. Accept the watermark or upgrade.
• You need a finished training or sales video without touching code: an avatar platform such as HeyGen gives you the editor, avatars and voices in one place.
• You have real footage and new audio, often in another language: use a re-sync or dubbing service that edits only the mouth.
• You want privacy and have a GPU: try an open-source local tool and budget setup time.
• You are building an app, running batches, or automating with agents: use an API. With Unicorn API you send an image and audio to OmniHuman 1.5, generate cutaways from the image-to-video models on the same key, and know the price of each request before it runs.

Whatever you choose, test with a 5 to 10 second clip of your real face and real audio before committing to a long video. Short tests reveal angle, lighting and audio problems for a few credits, and fixing those inputs does more for quality than switching between AI lip sync tools.