Creating AI images for videos: format, scenes and style

To create AI images for videos, it is not enough for each image to look good on its own: they need to come out in the video's format (landscape 16:9 for YouTube), with plenty of resolution so they can be moved, one per scene based on what the narration says, and all in the same style. Text and captions are added afterwards, in the edit, and if the result looks real, YouTube asks you to disclose it when you upload the video.

What changes when the image is for a video

Three things change. The image is not seen on its own but in sequence, so what matters is that the whole is consistent. It is on screen for several seconds and almost always moving, so it needs room to be cropped without losing sharpness. And it accompanies a voice: it has to show what is being said at that moment, not a generic illustration of the topic.

That is why the work does not start in the image generator, but in the script. First you decide what is told and in what order; then, what is seen in each stretch. If the video is narrated and you do not appear on camera, the images are almost everything the viewer is looking at, and it is worth planning them the way you plan a storyboard.

Format: 16:9 and resolution with room to spare

YouTube's standard aspect ratio on computers is 16:9, according to its help. If you generate the images already in landscape 16:9, you will not have to crop them or fill the sides with bars. Gemini image models on Google Cloud support that ratio: the Gemini 3.1 Flash Image page includes 16:9, and also 9:16, 1:1 and others, among its supported aspect ratios.

Resolution depends on the model. According to its documentation, Gemini 3.1 Flash Image supports 512, 1K, 2K and 4K; Gemini 3 Pro Image, 1K, 2K and 4K; and Gemini 3.1 Flash Lite Image, only 1K. For video there is a practical reason not to fall short: if you are going to apply a zoom in or a pan, the software enlarges the image and keeps part of it, and an image that is only just big enough loses sharpness. We explain it in the Ken Burns effect.

One image per scene, based on the narration

The tidiest way to produce is to split the script into short scenes and write, for each one, what is seen. An example with a popular-science video about an island lighthouse:

Scene narration	Image description
“For decades, the lighthouse keeper climbed the tower's one hundred and twenty steps every night.”	Stone spiral staircase inside a lighthouse, seen from below, warm oil-lamp light, watercolour illustration.
“From the top, the light reached the ships approaching the coast.”	Lighthouse on a cliff at night, beam of light over the sea, a sailing boat in the distance, same watercolour illustration.
“Today the lighthouse runs by itself and the keeper's house is empty.”	Small white house next to the lighthouse at dusk, closed windows, no people, same watercolour illustration.

Notice that each description names something concrete (the staircase, the beam of light, the empty house) and repeats the style at the end. Writing those descriptions has its own technique: we cover it in prompts for AI images.

The same style throughout the video

It is the most visible mistake in videos made with AI: one scene looks like a photo, the next like a comic and the third like a painting. To avoid it:

• Decide the style once and write it down. A fixed sentence (“watercolour illustration, ochre and blue tones, warm light”) added to all the video's descriptions, and if possible to all the channel's videos.
• Keep the palette. If the channel has its own colours, mention them in the style.
• Be careful with recurring characters. The same character generated several times can come out different in each image. If it is not essential, tell the story with shots where the character does not need to be recognised (hands, silhouettes, objects).
• Choose a level of realism and do not change it. Mixing photorealism with illustration is disorienting and also changes what you will have to disclose on YouTube (see below).

What not to ask the AI for

• Text inside the image. Captions, figures and titles come out cleaner and free of spelling mistakes if they are composed in the edit, with a real typeface. Ask for the image without text.
• Real people and celebrities. Besides the problem of imitating someone, Google's safety filters detect a photorealistic depiction of a celebrity in the request. And if the image makes it look as though a real person is doing or saying something they have not done, YouTube asks you to disclose it.
• Brands and logos. They add nothing to an illustration and can bring problems.
• Content the model will not generate. If an instruction is potentially unsafe, the model may refuse to generate the image or the filters may block it. Rephrase the scene instead of insisting.

What rights you have over what the AI generates, and what happens if an image looks too much like someone else's work, is a separate question: we deal with it in AI images and copyright.

Review, redo and disclose

Google's documentation puts it bluntly: do not expect a perfect image on the first attempt. Review each image with its narration next to it and redo the ones that fail. Look at hands and faces, at objects that do not fit the period, at text that appears even though you did not ask for it, and at details that contradict what the voice says.

Before publishing, decide whether the video needs the AI disclosure. YouTube asks creators to tell the audience when they use AI to significantly alter or generate photorealistic content, for example if realistic scenes that did not actually happen have been generated. By contrast, you do not need to disclose it if unrealistic content has been created with AI. A video illustrated with obvious watercolours is not the same as a photographic recreation of an event. The cases are in altered or synthetic content on YouTube.

How the Estudio de Canales does it

In the Estudio de Canales, images are a production phase, scene by scene:

• The description comes from the storyboard. The AI suggests, for each scene, what is seen and with what movement. Its instructions ask it to make the description follow the bible's audiovisual style, with no brands, no recognisable real people or celebrities and no logos. You edit it if you are not happy with it.
• Fixed channel style and format. The image style from the channel bible is added to each description, and the model is asked for a landscape 16:9 image, with no text, no lettering, no logos and no watermarks.
• Gemini image models, your choice. Gemini 3.1 Flash Image (2K), the recommended one; Gemini 3.1 Flash Lite Image (1K), the most economical; or Gemini 3 Pro Image (2K), the highest quality.
• All at once or one by one. You can request the images for the whole video at once or redo just one scene's; the others do not change. If you request again the image for a scene that already has a current one with the same description and the same model, the studio returns the existing one instead of generating it and paying for it again; redoing it on purpose is a new generation.
• If you change the description, the image is no longer valid and the scene's review asks for it to be generated again.
• Your own images fit too. You can upload an asset of your own with its source and licence, or reuse an image that is already in your channel.
• Provenance recorded and AI disclosure. Each generated image is marked as made with AI, with the provider and the model. If the video includes AI assets, the package for uploading comes with the synthetic content disclosure ticked; the final decision when uploading is yours.
• A person approves each scene before editing.

What it does not do: it does not generate video clips with AI (the scenes are images with movement), it does not search for photos in image libraries and it does not put text inside generated images; captions are composed in the edit. Images always come out in landscape, because the studio currently produces 16:9 video, not Shorts or vertical video.

Sources consulted

Checked on 11 October 2026. We link to the original source so that you can consult the current text.

• Configuración de codificación de subidas recomendada. Google · Ayuda de YouTube.
• Gemini 3.1 Flash Image. Google Cloud · documentación.
• Gemini 3 Pro Image. Google Cloud · documentación.
• Gemini 3.1 Flash Lite Image. Google Cloud · documentación.
• Prácticas recomendadas para la generación de imágenes de Gemini. Google Cloud · documentación.
• Generación de imágenes con Gemini y la IA responsable. Google Cloud · documentación.
• Avisar sobre el uso de contenido generado por IA. Google · Ayuda de YouTube.