# Talking photo AI: one line of dialogue in Veo 3.1, Kling and Seedance — tested on one portrait

> A talking photo takes one generation: upload a portrait, write the line in quotes, and the model animates the face, syncs the lips and voices the text itself. We gave Veo 3.1, Kling 3.0 and Seedance 2.5 the same frame and the same line and compared articulation, pauses and likeness. Below: clips with sound, the full prompt and how to make a talking avatar with your own voice.

Published 2026-10-03 · Updated 2026-10-03
URL: https://ai2face.com/blog/govoryashchee-foto

A talking photo is a video in which the person from a portrait looks into the camera and says a line of text: the lips move with the words, the eyes blink, the head moves slightly. This used to take two tools — one to animate the face, another to voice the text. In 2026, video models do both in one generation. We tested three of them on one portrait and one line in Russian.

## 1. How talking photo AI works

The photo becomes the first frame of the clip. The model generates the following frames: lip movement, facial expressions, blinking, small head turns. A line in quotes in the prompt becomes speech — the model generates a voice and syncs the lips to it. That's how [Veo 3.1](/models/veo-3-1), [Kling 3.0](/models/kling) and [Seedance 2.5](/models/seedance) work: all three create sound together with the video.

There's a second route too: upload a ready-made audio recording. Seedance 2.5 accepts up to 10 audio files as references, so the speech can be in your own voice or one recorded in advance.

## 2. How to make a talking photo: 4 steps

1. **Prepare a portrait.** The face large and facing the camera, mouth closed, even light. Pick the format for your platform: 9:16 for stories and Reels, 16:9 for YouTube and presentations.
2. **Write the line.** Up to 12–15 words for 8 seconds. Punctuation sets the pauses, so put periods and commas where a person would take a breath.
3. **Describe the voice and behavior.** Language, timbre, tone, facial expression: “speaks in a warm female voice, blinks, smiles at the end”.
4. **Rule out extras.** “Static camera, no music, no subtitles, no text on screen” — otherwise the model may add background music or captions.

## 3. Which photo works for a talking avatar

We took the portrait of a woman from our guide to [AI profile picture prompts](/blog/promty-dlya-avatarki) and turned it into a vertical “blogger at a desk” frame in [Nano Banana Pro](/models/nano-banana-pro).

![Vertical portrait of a woman in a green sweater at a desk in a bright home office, calmly looking into the camera](https://ai2face.com/_astro/start-frame.BGPN2OdS_26I8rG.jpg "The first frame for all clips: a large face, closed mouth, soft window light. Made in Nano Banana Pro from a selfie; the person was generated for this test")

What matters in the frame:

- **A closed mouth.** The model opens it on the first syllable; an open mouth in the photo causes a jump at the start of the clip.
- **The face in the top half of the frame.** In a vertical video that leaves room for the platform's interface.
- **Nothing near the face.** Hands, a microphone or hair over the lips interfere with lip sync.
- **Your own photo, or with consent.** A talking video of someone else's face without permission is a deepfake, not an avatar.

## 4. Test: one line in three AI models

One frame, one prompt, 8 seconds, 9:16. The line, in Russian: «Привет! Это обычное фото, а теперь оно говорит. Сделать так можно за минуту.» (“Hi! This is an ordinary photo, and now it talks. You can do this in a minute.”)

```
The woman looks straight into the camera and speaks in Russian in a warm,
natural female voice: «Привет! Это обычное фото, а теперь оно говорит.
Сделать так можно за минуту.» Accurate lip sync, natural blinking, small
head nods and a soft smile at the end, static camera, quiet room
ambience, no music, no subtitles, no text on screen
```

[A woman in a green sweater talks to the camera with lively articulation and nods](https://ai2face.com/_astro/media/blog/govoryashchee-foto/veo.mp4 "Veo 3.1, 9:16, 8 s. The most expressive face; starts speaking from the first frame")

[A woman in a green sweater calmly talks to the camera and smiles at the end](https://ai2face.com/_astro/media/blog/govoryashchee-foto/kling.mp4 "Kling 3.0, 9:16, 8 s. Starts with a closed mouth, speech takes about 5 seconds, then a pause and a smile")

[A woman in a green sweater talks to the camera with restrained articulation and pauses between phrases](https://ai2face.com/_astro/media/blog/govoryashchee-foto/seedance.mp4 "Seedance 2.5, 9:16, 8 s. Four phrases with clear pauses — exactly matching the punctuation of the line")

We compared the clips frame by frame and by the audio track:

| | Veo 3.1 | Kling 3.0 | Seedance 2.5 |
|---|---|---|---|
| Face and features | preserved | preserved | preserved |
| Start of the clip | speaks from the first frame | 0.5 s of silence, mouth closed | 0.6 s of silence, mouth closed |
| Pauses in speech | no drops into silence | one pause in the middle | four phrases, pauses of 0.4–0.9 s |
| Facial expression | the most expressive | calm, presenter-style | restrained articulation |
| End of the clip | speech almost to the end | 2 s of silence and a smile | 0.9 s of silence |
| Length per generation | up to 8 s | up to 15 s | up to 30 s |

Our takeaway: all three models kept the face, freckles and earring, and voiced the Russian line. Veo 3.1 looks the most alive — good for emotional social media clips. Kling 3.0 gives a smooth, calm clip with a pause at the end — handy for stitching several lines together. Seedance 2.5 follows punctuation most precisely and is the only one of the three that accepts your own audio recording. Listen to the pronunciation of any line before publishing: Google doesn't publish the list of languages in which Veo 3.1 voices speech.

## 5. How to make a photo talk in your own voice

If the voice must be yours, you don't need to generate speech — you can hand it to the model.

1. Record the line on your phone in a quiet room, 5–15 seconds.
2. Upload the portrait as the first frame and the audio file as a reference in Seedance 2.5.
3. In the prompt, state that the speech comes from the audio, and the model handles only the lips, expressions and blinking.

The same method works for a voice-over by a narrator or for a song: the model syncs the lips to any uploaded track.

## 6. A talking cartoon avatar

It's not only photos that can talk — an illustrated character can too. We took a 3D avatar from our prompts guide and gave Kling 3.0 a short line in Russian: «Всем привет! Я — ваш новый аватар.» (“Hi everyone! I'm your new avatar.”)

[A 3D cartoon woman with freckles on a teal background talks to the camera and winks](https://ai2face.com/_astro/media/blog/govoryashchee-foto/cartoon.mp4 "Kling 3.0, 1:1, 5 s. Cartoon-style articulation and a wink. Flaw: at second 4 an extra hand is visible at the left edge of the frame")

Models get away with more on cartoon characters: the lips are stylized anyway, so small sync errors go unnoticed. Still, check the whole frame — in our clip a hand appeared at the edge that wasn't in the prompt. How to make the avatar itself is covered in our [AI profile picture generator](/blog/avatarka-nejroset) guide.

## 7. What talking photos are good for

- **Social media and stories.** A short greeting, an announcement, a reply to followers without filming.
- **Training and presentations.** A host character explains a lesson step or a slide.
- **Ads and product listings.** A person with the product says one line about its main benefit.
- **A blog avatar.** The same character in every video — no camera or lighting needed.

For clips where a photo should simply come alive without speech — smile, blink, turn the head — see our guide on how to [animate a photo with AI](/blog/ozhivit-foto).

## 8. Free talking photo AI: what's realistic

Video with sound is the most expensive kind of generation: models charge for every second of the clip. There are no fully free services at the quality level of Veo 3.1. AI2Face gives you starter credits when you sign up, and the cost of each clip is shown before you run it. To spend less, fine-tune the line on 4–5 seconds and only then make the final version.

## 9. Limitations and ethics

- **Length.** Per generation: up to 8 seconds in Veo 3.1, up to 15 in Kling 3.0, up to 30 in Seedance 2.5. Split long speech into clips.
- **Pronunciation.** Models sometimes misplace stress or misread rare words — listen to every clip before publishing. Spell tricky words the way they sound.
- **Extra details.** Models may add hands, objects or camera movement — describe in the prompt what shouldn't be there.
- **Other people's faces.** Don't make talking videos of people who haven't agreed to it, and don't put words in real people's mouths that they never said.
- **Photos of relatives.** Animating a family archive can be moving, but the decision belongs to the family; a talking video of someone who has passed away may hurt loved ones.

## Bottom line

In 2026 a talking photo takes one generation: the portrait as the first frame, the line in quotes, a description of the voice — and the model voices the text and syncs the lips itself. In our test all three models voiced a Russian line while keeping the face. Use Veo 3.1 for lively, emotional clips up to 8 seconds, Kling 3.0 for a calm delivery and stitching lines together, and Seedance 2.5 for clips up to 30 seconds and voicing with your own recording. Before publishing, listen to every clip and check the whole frame. In AI2Face all three models are in one generator, and the cost of each clip is shown before you run it.

## FAQ

### How do I make a photo talk?

Upload a portrait as the first frame of a video and write the line in quotes in the prompt, along with the language and voice: “speaks in a warm female voice: ‘…’”. Veo 3.1, Kling 3.0 and Seedance 2.5 voice the text and sync the lips in one generation.

### Which AI makes talking photos in languages other than English?

In our test all three models — Veo 3.1, Kling 3.0 and Seedance 2.5 — voiced a line in Russian. Veo 3.1 had the most expressive face, Seedance 2.5 the clearest pauses between phrases, and Kling 3.0 the calmest, presenter-style clip.

### Can I make a photo talk in my own voice?

Yes. Seedance 2.5 accepts up to 10 audio files as references — upload a recording of your voice and state in the prompt that the speech comes from it. Without audio, the models generate a voice from your description.

### Can I make a talking photo for free?

Partly. AI2Face gives you starter credits when you sign up. Video with sound costs more than images, so start with a 4–5 second clip. The cost is shown before you run it.

### How long can a talking video be?

Per generation, Veo 3.1 makes up to 8 seconds, Kling 3.0 up to 15 and Seedance 2.5 up to 30. Eight seconds fits a sentence of 12–15 words; split longer text into several clips.

## Sources

- https://ai.google.dev/gemini-api/docs/veo
- https://kling.ai/quickstart/klingai-video-3-omni-model-user-guide
- https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5