How to Make a Photo Sing With AI in 2026: Step-by-Step Guide

To make a photo sing with AI, you need two steps: first a song, then a portrait that "performs" it. You can generate a track with your own lyrics in ElevenLabs Music (30 seconds to 3 minutes), then upload the photo and the audio to OmniHuman 2.0 or Kling Avatars v2. The model returns a video in which the person, character or pet in the picture sings with matching lip movement, facial expression and swaying to the beat, and the clip is as long as the song, up to 5 minutes. It all runs in the IIshki studio without a VPN or a foreign card. Below: which model to choose, a step-by-step walkthrough and the mistakes that make lips miss the words.

What you need for a singing photo

Three things:

1. A photo: one portrait with a clearly visible face. A phone shot, an old scanned print, a drawing or a picture of your pet will all work.
2. A song: your own recording, a track you hold the rights to, or a song generated by AI from your own lyrics.
3. An avatar model: it "fits" the sound onto the face, moving the lips to the words and adding blinks, head tilts and expression.

If you want the photo to talk rather than sing (a greeting, an announcement, a voice-over), see our separate guide to the talking photo. Here we focus on songs: long vowels, instrumental breaks and rhythm make models behave differently than with speech.

Which model to choose

To compare other lip-sync tools, see our roundup of AI lip-sync tools.

How to make a photo sing: 5 steps

1

1. Prepare the photo

A front-facing or slightly turned portrait where the face takes up a good part of the frame works best. The requirements are simple: at least 300 pixels per side, JPG, PNG or WebP. The fewer distractions there are (hands near the face, glasses with glare, hair over the lips, a second person nearby), the more accurately the model hits the words. A faded old photo can be cleaned up first, and if you have no suitable shot, you can generate a character portrait in FLUX 3:

Studio portrait of a young woman with short hair, looking straight into the camera, mouth closed, neutral grey background, soft frontal light, waist up

2

2. Make or pick a song

If you have no recording, open ElevenLabs Music. Describe the sound in the description field and paste the words into the Lyrics field, marking the sections:

Description: warm acoustic pop ballad, female vocal, guitar and light drums, slow tempo

[Verse] I remember the town in the first snow And the light in a window across the shore

[Chorus] Sing with me while the fire still glows

Track length is 30 seconds, 1, 2 or 3 minutes; 30 seconds is enough for a first test. Don't put the names of real artists or bands in the description: the model will reject such a request. Our guide to a song with your own lyrics in ElevenLabs Music covers lyric markup and writing lines that sing well.

3

3. Check the audio against the model

Before uploading, compare the file with the model's limits:

• OmniHuman 2.0 MP3 or WAV, 4 seconds to 5 minutes, up to 15 MB.
• Kling Avatars v2 MP3, WAV, M4A or AAC, 2 seconds to 5 minutes, up to 5 MB.

A long, high-quality song may not fit into 5 MB. Then you have two options: upload it to OmniHuman, which has a higher limit, or cut out the chorus in any audio editor. For social media, a 15–40 second fragment often works better than the full song.

4

4. Upload the photo and the track

Open OmniHuman 2.0 or Kling Avatars v2, put the portrait in the photo slot and the song in the audio slot. Clip length is set automatically from the track. In Kling Avatars you can choose Std or Pro quality.

5

5. Tell it how to sing

A prompt is optional: the sound drives the lips. But a short hint changes the delivery: emotion, gaze, gestures. OmniHuman allows up to 300 characters for it.

Sings with eyes closed, sways gently to the beat, smiles and looks into the camera on the chorus

Sings a rock chorus with energy, nods to the rhythm, eyebrows raised, static camera

Run a short fragment first to check how the model holds the face and hits the words, and only then render the full song.

How to match the song to the photo

The avatar model moves the lips to the voice, so the track matters as much as the picture. A few rules that save reruns:

• A voice that fits the face. An older man in a vintage photo calls for a low male vocal, a child in a drawing for a light, high one. When the voice and the face disagree, the clip looks dubbed even if the lips hit the words. In ElevenLabs Music you set the voice type right in the description: "low male vocal", "bright children's choir", "husky female voice".
• Vocals in front of the music. If the voice drowns in the instruments, the lips have nothing to follow. Ask for "clear vocal up front", "minimal instruments", "acoustic" in the song description.
• Singing from the first seconds. A long intro without vocals means seconds where the character just stands there. For a first test, trim the start so the verse or chorus kicks in right away.
• Tempo. A slow ballad produces large, clearly visible mouth movements and is the easiest way to test a photo. Fast rap works too, but flaws are harder to spot, so start with something slow.
• Language. You can sing in English, Russian or another language: write the lyrics in ElevenLabs Music in that language, and the avatar model simply follows the sound.

If a pet or a cartoon character sings

OmniHuman 2.0 takes more than people: you can put a cat, a dog or a cartoon hero in the photo slot. For animals, a shot where the muzzle faces the camera and the mouth is clearly visible matters most. A side view, a muzzle in the grass or fur covering the mouth are common causes of "melting" expressions. For drawn characters, use a crisp drawing with a clearly drawn mouth: the model animates what it sees.

Common mistakes

What to do with the finished clip

• Raise the quality. If the original photo was small, run the finished video through the Topaz upscaler.
• Build a music video. Scenes for the instrumental breaks can be animated from photos; see our roundup of AI tools for photo animation.
• Put the whole chain on the canvas. Portrait, song and avatar connect into one flow on the canvas, so you can rerun any step on its own.

Start small: a front-facing portrait, a 30-second chorus from ElevenLabs Music and OmniHuman 2.0. Once the lips hit the words, render the whole song. The price of every generation is shown on the button before you start.