Using references

A reference shows the model what you want your result to look like. In the first video, we used a single image to go from image to video. Now let's look at everything references can do.

There are a few ways to add one: drag it in from your library, drop a file from your computer straight onto the area, or click an empty box and choose a file.

You can also right-click a file in your library and choose Use as Reference, or right-click an image box to paste an image you've copied.

Every image you upload goes through an automatic safety check, and every generation that uses an image is checked again.

The check looks for nudity and sexual content. If there's a face in the image, your prompt is read together with it.

Here's why: photos of real people can be misused to make fake images that undress or sexualise them, and those images are used for harassment and blackmail. Putting a face into a non-sexual scene is fine.

A blocked image isn't uploaded, and a Content blocked window opens; a blocked generation costs no credits. If you think it's a mistake, appeal from the same window, and someone on our team will review it and reply by email.

The Reference area changes shape with the model you pick. Each box is tagged image, video or audio, and the top shows how many references the model takes in total.

Models that take a first and last frame need two images. The video starts on the first one and ends on the last.

While a required box is empty, an amber warning shows underneath, and the cost appears once the reference is added. Click Generate at that point, and LUVI tells you which reference is missing.

With FLUX 3 Keyframes, you can add up to ten images and set the moment each one appears in the video.

Editing models like Nano Banana 2 take up to fourteen image references at once, and you can drag several files in from your library together.

Type @ in the prompt and pick a reference, and a chip goes into your text; hover over it to preview the image. That way, the model knows exactly which image you mean.

Talking-avatar models need a face image and an audio file. Drop different kinds of files in together, and each one lands in its own box.

With Auto models, your references also decide which version runs. On MiniMax H3, no reference means text to video and one image means image to video. Two images become a first and last frame, and more images, a video or audio switch it to reference to video.

The small icon on the model button always shows which version was picked.

To remove a reference, hover over it and click the X. To remove them all, use Clear All.

Switch to a model that takes fewer references, and LUVI tells you how many will be removed and asks you to confirm. The ones that fit move across by type.

Now let's watch the video made from the first and last frames.