How to Make a Photo Talk With AI: Photo, Audio, Settings

You make a photo talk by uploading a portrait and an audio track to the Avatar tab, choosing a lip-sync model, and generating. The portrait can be up to 10 MB and the audio up to 100 MB. Model choice is mostly about length: Omnihuman 1.5 takes audio up to 60 seconds, InfiniTalk and WAN speech-to-video take up to 120 seconds, and the Kling Avatar models accept audio up to 5 minutes. Run a short test before the full clip.

Updated October 2026.

What a talking photo is and what the Avatar tab needs

A talking photo is a still portrait animated so the mouth moves in time with an audio track. The head may move slightly and the eyes may blink, depending on the model, but the picture stays close to the original still. It is not a full-body animation and it is not a generated scene; it is a portrait plus a voice.

The Avatar tab takes two inputs: a photo and an audio file. The photo is uploaded from your device up to 10 MB, and the audio up to 100 MB. The audio length is what decides the model, because each model has a different cap. The tab also accepts a video input on the Video Lip Sync option, which replaces the photo with an existing clip.

There is no prompt to write for the model itself, so quality depends on the photo and the audio. If you are choosing between routes, which lip-sync tool to use compares them, and the talking avatar page covers the overall workflow.

Pick the photo

The photo does most of the work. A model can only animate what it can see, so the framing and the lighting decide how believable the result is.

• Front-facing. The face should look toward the camera, or close to it. A strong side profile hides one side of the mouth and gives the model less to work with.
• Mouth visible. No hand, microphone, scarf or hair across the lips. The mouth needs to be clear and evenly lit.
• Shoulders in frame. A headshot that includes the shoulders gives the animation somewhere to sit. A tight crop on the face can look like a floating mask when the head moves.
• No sunglasses. Eyes matter for believability, and dark lenses hide them.
• Even light. A hard shadow across half the face makes the animation look patchy. Soft, frontal light is best.
• Neutral expression. A wide open-mouth smile is harder to animate than a relaxed, closed-mouth expression.

Use the highest resolution you have. A small, compressed image gives the model less detail to work with, and the result will look softer than the source. For more on preparing an image, see writing a better AI image prompt.

Get the audio

There are three practical routes.

1. Record it. A phone recording in a quiet room is fine. Speak a little slower than usual, keep a steady distance from the microphone, and leave a beat at the start and end.
2. Upload an existing track. A voice memo, a podcast segment or a narration file works, provided you have the right to use it.
3. Generate the voice. Write the script, then use text to speech to produce the audio. That is the cleanest route when the photo is of a persona rather than a real recording session.

Whatever the source, listen to the whole track before uploading. The model follows the audio it is given, so a track with a loud room hum or a clipped peak will carry that problem into the video. Trim long silences at the start and end, because they add length without adding anything to the clip.

Choose a model by length

The model list is organised by how much audio each one accepts and what it costs. Prices are per second of audio, with a minimum on some models.

Model	Audio length	Price	Notes
Omnihuman 1.5	Up to 60 seconds	10.8 credits per second	High-quality talking avatar; the per-second rate applies to audio length
InfiniTalk From Audio	Up to 120 seconds	1.2 credits per second at 480p, 4.8 at 720p	Minimum 15 credits
WAN Speech To Video Turbo	Up to 120 seconds	4.8 credits per second at 480p, 9.6 at 720p	Minimum 15 credits
Kling Avatar Standard	Up to 5 minutes	954 credits per 300 seconds	720p
Kling Avatar Pro	Up to 5 minutes	1,910 credits per 300 seconds	1080p
Video Lip Sync	Video input plus audio	3.2 credits per second	Lip-syncs an existing video to new audio

There is also a lighter lip-sync model at 1.1 credits per second with a 5-credit minimum, which suits a short test where cost matters more than the last bit of realism. The price for the exact clip is shown before you generate, and a failed generation is refunded.

For a 20-second message, Omnihuman 1.5 is the quality option and InfiniTalk at 480p is the cheap one. For a five-minute lesson, the Kling Avatar models and PixVerse lip-sync are the ones that accept the length. Match the model to the audio before you upload, not after.

Run a short test, then the full clip, and fix what lip sync gets wrong

Generate a 5 to 10 second test from the same photo and the same voice before committing to a long render. A short test shows whether the framing works, whether the mouth is visible, and whether the voice fits that face.

When the result is not convincing, the cause is usually one of these:

• Fast speech. Very fast delivery gives the mouth more shapes per second than the model can resolve. Slow the script down or split it into shorter takes.
• Side profile. A face turned away from the camera gives the model too little mouth detail. Use a more frontal photo.
• Teeth and lips. A very open smile or a hand near the mouth confuses the animation. A neutral, closed-mouth source is more reliable.
• Audio with music. A track with a loud music bed under the voice makes the timing less clear. Use a clean voice track for the lip-sync pass and add music later in the timeline.

These are the failure modes to look for, not a guarantee that every clip will have them. Watch the first few seconds closely, then the end; problems often show up at the edges of a long track.

Consent: only photos you have the right to use

This is the line that matters. Use a photo of yourself, or a photo of someone who has given you clear permission to animate their face and voice. Do not animate a photo of a public figure, a colleague, a customer or anyone else without their consent, and do not use a real person’s face to say words they did not say.

When the result looks like a real person, label it as synthetic where the platform or the jurisdiction requires it. Advertising platforms increasingly expect viewers to know when a person on screen is generated. Where a rule applies, follow it, and check the rules for your market rather than assuming one standard covers everywhere.

If the photo is of a child, get consent from a parent or guardian. If the audio is someone else’s voice, get their permission too. The voice and lip-sync guide covers the disclosure side, and the AI avatar generators round-up compares the wider tool set.

Frequently asked questions

How long can the audio be?

It depends on the model: 60 seconds for Omnihuman 1.5, 120 seconds for InfiniTalk and WAN speech-to-video, and up to 5 minutes for the Kling Avatar models. Match the audio length to the model before uploading.

What photo should I use?

A front-facing portrait with the mouth visible, shoulders in frame, no sunglasses and even lighting. A neutral expression animates more reliably than a wide open-mouth smile.

Can I use a photo of someone else?

Only with their clear consent. Do not animate a public figure or anyone else’s face without permission, and label synthetic results where required.

Can I use the result commercially?

Check the terms of the model you used and the rules of the platform where you will publish. Do not assume a generated clip is cleared for advertising without reading them.

Which languages work?

The models follow the audio they are given, so the language is set by the audio track rather than by a menu in the tab. Generate the voice in the language you need and upload that track.

Open the Avatar tab, upload a front-facing portrait and a short audio track, and generate a test with the model whose length fits. New accounts start with 10 free credits and no card required, so you can see how your own photo handles the first few seconds before committing to a full clip. If you need a voice first, generate one in text to speech; credit packs are one-time purchases that never expire, and the options are on the pricing page.