What a talking photo is
A talking photo starts with two inputs: one still image of a person and one audio track. An AI video model reads the sound wave of the voice, maps the phonemes to mouth shapes, and generates new video frames where the person in the photo speaks the line.
Older tools only pasted a moving mouth onto a frozen face. The jaw moved, the cheeks stayed still, and the eyes stared at nothing. Current audio-driven video models animate the whole frame. When the voice takes a breath, the chest and shoulders rise. When a word hits a higher pitch, the eyebrows lift and the head tilts. The eyelids blink between phrases, and the hands move if they are visible in the frame.
In Doppelshot, you do not record a reference video or upload ten angles of your face. You open the Say it card, upload one selfie or pick an AI portrait you already made in the app, type the lines you want spoken, and choose a voice. The app generates the speech track first, then passes the photo and the audio to the video model. You get a vertical video ready for TikTok, Instagram Reels, or YouTube Shorts.
Pick the photo
The video model can only work with what is visible in the starting image. Every hidden feature forces the AI to guess, and guesses are where faces start to drift.
Use a front-facing photo where the person looks straight at the lens or slightly off-center. The mouth must be clearly visible. A relaxed, closed mouth or a slight smile gives the cleanest starting point. If the person is laughing with a wide open mouth, the model treats that open jaw as the resting face, which makes the pauses between words look strange.
Keep the face clear of obstacles:
- No hands covering the chin or cheek.
- No microphones, coffee cups, or food in front of the lips.
- No heavy shadows cutting across the mouth or eyes.
- No sunglasses, because eye contact does half the work of making speech look real.
You can use a raw phone selfie, a studio portrait, or a stylized image generated on a Doppelshot photo card like a vintage film portrait or a movie poster. Both close-ups and waist-up shots work. A close-up focuses attention on the eyes and lip sync. A waist-up shot gives the model room to move the shoulders and hands while speaking. Every person in an uploaded photo must be an adult (18+) who has agreed to appear in AI-generated content.
Write for nine seconds
Most short talking clips run for nine or ten seconds. In normal human conversation, nine seconds is about 20 to 25 words.
If you paste a 60-word paragraph into a short clip, the speech generator either rushes through the text or gets cut off mid-sentence when the video ends. If you write three words, the person speaks for one second and spends the remaining eight seconds blinking in silence.
Write the way people talk, not the way they write emails:
- Use short sentences. Two sentences of ten words each sound much more natural than one twenty-word sentence with three clauses.
- Use punctuation to control the breathing. A comma gives the voice a brief pause. A full stop makes the speaker close their mouth, drop their chin slightly, and take a breath before the next line.
- Write numbers and abbreviations out in full words so the voice reads them the way you intend.
- Read your script out loud at a normal, unhurried pace while watching a timer. If you pass nine seconds, cut a few words.
When you need a longer monologue or an announcement, switch to a model that supports extended durations, or build a two-person dialogue on the Make a podcast card.
The voice: pick one or clone yours
The realism of a talking photo depends as much on the audio track as on the video pixels. Because the video model takes its cues from the sound wave, a flat, robotic voice produces a flat, stiff face.
Doppelshot uses ElevenLabs (Eleven v4) for its voice library. You can pick from about 30 voices across styles and tones, or leave the selector on Auto so the app picks a voice that matches the person in your photo. The voice speaks whatever language you write your script in. You can also give the voice a delivery note, such as soft, amused, or out of breath, so the line sounds spoken rather than read off a page.
If the photo is of you, a stranger's voice coming out of your face can feel odd to people who know you. You can clone your own voice directly in the app. You record a short consent clip once in your browser, and the clone is ready in a few minutes. It sits next to the library voices and works on every talking card.
Give it a direction
A person speaking on camera rarely freezes in one posture from the first word to the last. They look up while thinking, nod on an important word, or smile after delivering a punchline.
Several talking models let you add a short text direction alongside your script. This direction tells the video model what the person and the camera are doing while the audio plays.
Keep the direction simple and physical. Describe one or two small actions that fit inside nine seconds:
- "Looks straight into the lens, nods once, then smiles at the end."
- "Talks calmly with a slight head tilt, warm indoor light."
- "Holds a coffee mug, looks up at the camera and speaks softly."
Avoid stacking multiple big actions into a short clip. Asking a person to stand up, walk across a room, wave, and sit back down in nine seconds forces fast, unnatural motion that pulls focus away from the face and breaks the lip sync.
Which model: OmniHuman, LTX, HeyGen, Kling, Creatify
Each talking card in Doppelshot runs on a default engine, and you can switch to another model chip before you generate. Different models are built for different shots. The price shows before you run.
OmniHuman 2.0 (ByteDance). Built for expressive speech, body language, and singing. It tracks the voice closely, matching the lips, breath, and hand gestures to the audio. Because it takes a written direction, you can tell the person how to react while speaking. It works on close-ups and full-body shots, and runs in 10, 20, or 30-second lengths.
LTX 2.5 (Lightricks). Built for a cinematic shot up to 9 seconds (about 25 words). While most avatar engines focus strictly on the subject, LTX treats the clip like a film director: it animates the talking person inside a complete scene with framing, lighting, camera motion, and room tone mixed in after. Pick it when the environment around the speaker matters as much as the words.
HeyGen Avatar 4. Tuned for facial expression, brow movement, and eye contact. It keeps the look of the original photo while giving the speaker the presence of an on-camera host. It also powers the Three emotions card, where your photo delivers the same line three ways in one clip.
Kling Avatar and Kling Avatar Pro (Kuaishou). The practical choice for long scripts. Kling Avatar handles talking videos from 10 up to 60 seconds, priced per 10-second segment. The standard tier is the lowest-cost way to run a long announcement or intro; the Pro tier keeps finer facial detail and steadier motion over long takes.
Creatify Aurora. Trained on presenter and user-generated ad footage. It gives the speaker a direct, confident address to the camera, built for product pitches and social video hooks up to 20 seconds.
For step-by-step instructions on setting up your first clip, see the help page on how to make a photo talk or sing.
Where it fails, and the fix
Even good models fail when the inputs push them outside what they were trained to do. Here are the most common problems and how to fix them before you spend credits.
The face warps when the head turns. You uploaded a side profile or a three-quarter angle where only one eye and half the mouth are visible. When the speaker turns toward the camera, the AI has to invent the missing half of the face. Fix: use a front-facing photo where both eyes and both corners of the mouth are clear.
The teeth or lips look blurry. The original photo is low resolution, heavily compressed, or dark around the mouth. Fix: start with a well-lit selfie, or run the photo through the upscaler tool first so the model has sharp edges for the lips.
The speech gets cut off at the end. The script has too many words for the clip length you selected. Fix: trim the script to around 20 to 25 words per 10 seconds, or switch the length selector to 20 or 30 seconds.
The voice pronounces a name or acronym wrong. Speech models guess phonetic spelling from the language of the text. Fix: spell the tricky word phonetically in the script box so the audio generator reads the exact sounds you want.
The person looks stiff between sentences. There is a long silence in the script or no direction given. Fix: remove empty pauses by tightening the text, or switch to OmniHuman and add a gentle action note like "smiles slightly at the end."
FAQ
Can I make a photo talk in my own voice?
Yes. You record a short voice sample once in the Doppelshot browser app. Your cloned voice is ready in minutes and appears next to the 30 library voices on every talking card.
How long can an AI talking photo video be?
Up to 60 seconds, depending on the model you choose. LTX 2.5 makes clips up to 9 seconds, Creatify Aurora goes up to 20 seconds, OmniHuman runs up to 30 seconds, and Kling Avatar supports scripts from 10 to 60 seconds.
Do I need to upload 20 selfies to make a talking video?
No. One clear, front-facing photo is enough for every talking and singing card. You only use 10 to 20 selfies if you want to create a reusable AI twin for generating new photo scenes first.
Can I use the talking video for commercial projects or social media ads?
Yes. Every paid result on Doppelshot includes a commercial license with no visible watermark, and carries a machine-readable AI marker in line with EU rules. Free preview results carry a visible mark and are for personal use only.
