← ABUZ8 BLOG

AI Lipsync: Make Any Photo Talk With One Audio File

An AI lipsync tool takes a still photo and an audio file and produces a video of the person in the photo speaking the audio. HeyGen built a $500M company doing exactly this. The open-source models — SadTalker, Wav2Lip, MultiTalk, and now LatentSync — match HeyGen's quality on most prompts and run free on your own GPU. This post is the model stack, the use cases that matter, and the rules around using this tech without crossing into deepfake territory.

Our free lipsync tool wraps the open-source pipeline below.

What "lipsync" actually means in 2026

Three increasingly powerful flavors:

For most use cases (UGC, explainers, talking-product demos), tier 2 is plenty.

The model we use

Inside ABUZ8 the avatar pipeline is built on SadTalker for stills, MultiTalk for higher-fidelity audio-driven animation, and a CodeFormer pass at the end to clean up edge artifacts. Stack details:

• Input: one front-facing photo (any face, any age) + audio file (any length, any voice).
• SadTalker generates the head animation at 30fps from the audio's prosody.
• MultiTalk handles the precise lip-shape matching frame-by-frame.
• CodeFormer smooths the boundary between the animated face and the static body/background.
• Output: MP4 at the source photo's resolution, audio synced.

Render time on a modern high-end GPU: about 30 seconds per spoken minute. On a mid-range card: closer to 90 seconds per minute.

What it's actually good for

• Multi-language explainers. Record once, generate the video in 20 languages from translated audio. Used to require 20 separate shoots.
• Faceless brands giving themselves a face. Fictional spokesperson, consistent across every video.
• UGC at scale. Stock photo + script + voice = 100 personalized testimonial-style videos. (Disclose this. See ethics section.)
• Onboarding videos. Personalize "Welcome, $NAME" videos at scale without re-recording.
• Historical figures saying their own writings. Education use cases. Lincoln reading the Gettysburg Address from a photo and TTS audio.

What it's bad at

• Profile shots. The models are trained on front-facing portraits. A 3/4 angle photo produces unnatural mouth deformation.
• Tiny faces in the source. If the face is less than 200x200 pixels in the source photo, the output looks blurry. Crop tighter or upscale first.
• Heavy beards or face coverings. The model has trouble inferring the mouth shape under a beard. Output mouth often looks "behind" the beard.
• Singing. Trained on speech, not vocals. Singing audio produces weird mouth shapes.

The ethics that actually matter

This is the section every honest article on this topic needs. AI lipsync is dual-use technology. Same tool that creates a multilingual explainer also creates a fake video of a real person saying things they didn't say. The line:

• Lipsyncing your own face with your own voice (or voice-cloned with consent): clearly fine.
• Lipsyncing a fictional character or stock model for marketing: fine, label it as AI-generated where required (FTC guidance, EU AI Act).
• Lipsyncing a public figure in a way that could be mistaken for them actually speaking: not fine. This is the deepfake category.
• Lipsyncing a real private person without consent: not fine, possibly illegal depending on jurisdiction.

Legitimate platforms in this space (HeyGen, Synthesia) require consent verification before you can use a real person's likeness. Open-source tools have no such guardrails. Use the tech responsibly or it gets regulated into uselessness for everyone.

The workflow that produces broadcast-quality output

1. Source photo: high-res, front-facing, neutral expression, soft lighting. Avoid harsh shadows on the face.
2. Audio: clean, denoised, normalized to -3dB. Background hiss in the audio shows up as twitchy mouth movement.
3. First pass: render at preview quality (640px, 24fps). Check timing and expression.
4. Final pass: render at full quality (1080p, 30fps).
5. Post: light color grading and a subtle drop shadow if the face is composited onto a background.

The whole thing takes 5 minutes of human time and 30-60 seconds of GPU time. HeyGen charges $24/month for this. Open-source costs your electricity.

Try the free tool

The ABUZ8 lipsync tool takes one photo + one audio file and returns the talking-head MP4. SadTalker + MultiTalk + CodeFormer pipeline, no watermark. Free for clips under 30 seconds.