How to keep the same character across videos and images

The short answer

To keep the same character across several clips, create one clear image of the character once and use it in every clip: as a reference image in the "References" mode, or as a start image in the "Image to video" mode. Alongside that, write the same description of the character in every prompt, word for word, and keep the same model and the same settings.

How many reference images you can attach to a clip here: up to 30 in Seedance 2.5, up to 10 in Wan 3.0, up to 9 in Seedance 2.0 and MiniMax H3, up to 7 in Kling O3 and up to 3 in Veo 3.1, in an 8 second clip. For images, Google presents Nano Banana Pro as a model built to keep the look of up to 5 people.

For example, a 5 second clip at 720p with reference images in Kling O3 costs ₪1.9719.69 crKling O3, 720p · 5s · no sound right now.

Three ways to keep a character

The important difference: a start image sets the first frame, while a reference image only shows the model what the character looks like, and the model builds a new scene around it. In most models you choose one of the two ways in a generation, not both together. In Kling O3 you can combine a start image with reference images here.

A reference video is useful when you already have a clip of the character. According to Alibaba, in Wan 2.6 a clip of a character keeps its look and its voice in new scenes. In Kling O3 a reference video is used to create a new shot, such as the next shot, and according to Kling the result is a new clip, not an edit of the clip you uploaded.

How many reference images each video model takes

The table shows what you can attach here in each of the video models, and what the maker itself says about keeping the character.

Seedance 2.0 Mini also takes up to 9 reference images here, at 480p or 720p, so you can use it to test at a low resolution how the character comes out before moving to a bigger model. Its price is in the table at the end of this guide.

For Veo 3.1 Lite Google documents no reference images, but here you can pick up to 3 reference images in it in an 8 second clip, and in our test two reference images held through the whole clip. It is the low cost way to keep a character in the Veo family, and Veo 3.1 or Veo 3.1 Fast fit when the character also speaks.

The same face across several images

With images too, the way is to attach a picture of the character and ask for a new scene. The image models here take input images, so you can put the same character in another place, in other clothes or in another pose:

• Nano Banana Pro. Takes up to 14 images, and according to Google up to 6 of them of objects, up to 5 of characters and up to 3 of style. Google presents it as a model built to keep the look of up to 5 people, and its model card notes that keeping a character is not always perfect
• Nano Banana 2. According to Google’s documentation, it takes up to 14 reference images to keep characters and products consistent
• GPT Image 2. Takes up to 16 images in one edit, and OpenAI writes that every input image is processed at high fidelity
• Seedream 4.5. According to ByteDance it keeps the facial features, the lighting and the color tone of the reference image
• FLUX.1 Kontext [pro]. According to Black Forest Labs it keeps the same character across different scenes. Here it takes one input image per edit

And if the image also has Hebrew text, such as a product name or a campaign headline, in our experience the new GPT Image models are the best at it, and Nano Banana writes good Hebrew too. The full comparison of the two models is in the guide Nano Banana Pro vs GPT Image 2.

A workflow from one character to a whole series

This is how you build a series in which the character stays the same character, from the first image to the last clip:

1. Create a clear image of the character. In an image model, with the character in the center, simple lighting and a clean background. In models that take several images you can prepare several images of the same character, for example from the front and from the side
2. Write a fixed description of the character. One sentence in English with what matters to keep, such as age, hair and clothes, and copy it word for word into every prompt
3. Tell the model what each image stands for. In the example in Google’s guide for Veo 3.1, the prompt starts with the words Using the provided images for the detective, the woman, and the office setting, and only then describes the shot
4. Use the same images in every clip. The same reference images, or the same start image, throughout the series
5. Keep the same settings. The same model, the same resolution and the same aspect ratio, and the same words for style and lighting
6. Continue from the last frame. When a clip has to continue the previous one, save its last frame as an image and upload it as the start image of the next one
7. Fix only what drifted. A clip in which the character came out different is made again, and the rest stays as it is

Every rule for writing the prompt, including a glossary of camera moves, is in the AI video prompt guide.

The same product across a campaign

A product works exactly like a character, and sometimes it is even simpler. When the clip has to start from the real product, upload the product photo as the start image, and the first frame is the product itself. When the product has to appear in several different scenes, attach the photo as a reference image and describe the new scene in the prompt.

You can also go through an image model: attach the product photo to GPT Image 2 or Nano Banana Pro, ask for it in a new setting, and animate the image that comes out into a clip. The full explanation is in the guide how to turn an image into a video.

And the character’s voice

In a series, the voice matters as much as the look. Some video models can work from a voice sample or from a clip of the character:

• ByteDance writes that Seedance 2.5 keeps the looks and the voices of several characters through the clip, and here it also takes a sample audio file
• According to Alibaba, in Wan 2.6 and Wan 2.7 a reference video of a character keeps its voice too
• According to MiniMax, in MiniMax H3 references can guide the voice as well

Hebrew is a different story: no video model here officially supports Hebrew speech. So a steady Hebrew voice is created in a voice model whose official language list includes Hebrew, such as Eleven v3 or MiniMax Speech 2.8 HD, choosing the same voice from the voice list every time. You add the voiceover to the clip in editing. All the details are in the guide Hebrew AI voiceover.

What is still not perfect

Keeping a character has improved a lot, but the makers themselves note limits worth knowing:

• Google writes in the Nano Banana Pro model card that keeping a character is not always perfect
• ByteDance notes for Seedance 2.0 that lip sync can slip when several people talk in the same scene
• In Veo 3.1 reference images work only in an 8 second clip
• In Kling O3, you can attach up to 4 images together with a reference video
• In Seedance 2.5 a reference video is available here only at 480p or 720p

So check every clip before you publish, and make again only what did not come out well. And if a generation fails, all its credits come back to your balance automatically, as explained in the guide failed or stuck generations.

What it costs

The tables are computed as the page loads, with the same calculation the studio shows before every generation. Each row is one generation with a reference or input image, at the settings written in it.

Every model works from the same balance, so you can create the character image in an image model, the clips in a video model and the voiceover in a voice model, and pay only for what you send.

Questions and answers