Nano Banana Pro Character Consistency in Images and Video

An image model has no memory. Every run starts from noise and from the words you give it, and a description like a woman in her thirties with dark curly hair and a green coat fits a million faces. Run it ten times and you get ten people who match the sentence and not each other. No prompt fixes that.

A photo pins what a sentence cannot. When the model is handed an image and an instruction, the face, the hair line, the eyebrows and the way the coat sits all come from the pixels, and the instruction changes only what it names. That is why the method is edit first, generate second: the prompt describes the change, the photo carries the person.

Three things make a character readable across shots: the face, one wardrobe, and one or two props. Lock all three in the hero still and change at most one of them per shot. A new coat and a new haircut in the same image is a new person.

The sheet is six stills of the same character, each made by editing one hero image rather than by prompting from scratch. It takes a few minutes and gives every later shot a reference to start from.

The hero still

Open Nano Banana Pro and write the character as a photograph with a moment in it, not a passport picture: what they are doing, where, in what light. Name the wardrobe and one prop. Render at 2K so the face has detail to hold, and pick the still whose face you would recognise in a crowd; the whole method depends on it.

The edits

Attach the hero still as the photo to edit and give one instruction per run. Keep the character's description out of the instruction; the photo is the description.

For the profile row there is a second route: the Editor's Change Camera mode turns the camera around the same still to the left or the right. It takes no instruction at all, so nothing else is asked to change.

Save the six as the character's sheet. The plain background version is the one to animate and to edit next, because a model holds a face best when nothing competes with it. The Nano Banana page has more edit instructions written in this shape, and the image to image page walks through the edit step itself.

A photo edit is not a copy and paste. The model reads the photo, keeps what the instruction does not touch, and regenerates the rest. It holds some things far better than others, and knowing which saves a lot of re-rolls.

• Held well: the face's proportions, skin tone, hair colour and length, glasses, a beard, the colour and cut of clothing, the overall light direction.
• Held loosely: fine jewellery, logos, tattoos, patterns on fabric. Name them in the instruction if they matter.
• Regenerated: anything the instruction moves. Turn the head and the ears and the jaw line are redrawn.
• Not read at all: anything outside the frame. A full body edit from a head shot invents the body; make the full body shot early, from the widest hero you have.

Nano Banana 2 and Seedream 5.0 Pro edit a photo the same way and are the second opinion when Nano Banana Pro keeps missing an instruction. Nano Banana 2 costs half as much at 1K and 4K and about three quarters as much at 2K, which suits drafts; Seedream 5.0 Pro reads fabric and skin a little differently, which is sometimes what a stubborn shot needs. Try the same photo and instruction in Nano Banana 2 or Seedream 5.0 Pro before you rewrite the prompt.

Pick the single best shot from the sheet for each edit rather than hoping a set of photos will be averaged: the plain background hero for most, the full body shot for anything below the waist.

Three image models that edit a photo, and four video models that take the result, with the sizes, lengths and inputs the studio sells today.

• 
  GoogleNano Banana Pro1K, 2K, 4K
• 
  GoogleNano Banana 20.5K, 1K, 2K, 4K
• 
  ByteDanceSeedream 5.0 Pro1K, 2K
• KlingKling 3.03 to 15 s · 720p, 1080p
• KlingKling O3 Omni3 to 15 s · 720p, 1080p
• ByteDanceSeedance 2.0 Fast4 to 15 s · 480p, 720p

There are two ways to carry a character into a clip. A start frame says: here is the first frame, what happens next. Reference photos say: here is who this is, put them somewhere else.

The start frame route is the reliable one. The model begins on your still, so the first frame is the sheet, and the question is only how long the face holds; the shorter the clip, the less it drifts. Describe the motion only and keep the first second calm, as our image to video guide explains.

The reference route is the flexible one and the less exact one. Seedance 2.0 Fast reads the photos as a strong description of the subject, not as a locked frame, so the character appears in the new scene with the face and the outfit mostly right and the pose re-composed. Cite the photos in the prompt: @Image1 walks into the kitchen. Expect two or three runs before a keeper.

Kling Elements is a feature of Kling's own app for building a clip from photos of a subject (as of October 2026). DreamClip does not sell it: the Kling models here take a start frame and no reference photos, and the Kling page says what each one does.

A face is only a third of a character. Viewers recognise the coat and the bag before they study the eyes, so lock those in words as well as in the photo, and repeat the same words in every instruction and every motion prompt.

A fixed phrase for every prompt

Write one short phrase once and paste it everywhere: in the olive wool coat, with the brass pin on the lapel, hair tied back. Name the colour and the material, not the brand. Two props is the ceiling; a third is one the model will drop.

Light and lens

• Name the light once per scene and keep it for every shot in that scene: soft window light from the left, or hard noon sun.
• Keep one lens feel and one camera height per sequence. A wide shot and a tight portrait read as the same character; a fisheye does not.

When a scene changes the light, change it in the sheet first, then animate that still. A video model asked to take a daylight still into night is making two changes at once.

Here is the method on a short sequence: a courier, Lena, on one morning round. Every still is an edit of the sheet; shots 1 to 5 start from that still, and shot 6 brings Lena in through reference photos.

Every still names the coat and the pin, so even when a face softens in a wide shot the viewer still sees Lena, and shot 3 ends on Lena at the cafe door, the moment before shot 4, which hides the cut.

The expensive mistake is the first one, a weak hero still, because every later shot inherits it; spend the re-rolls there.

1. Write the hero prompt once and run it on up to four image models in Compare to pick the face you like. Each lane runs at that model's opening settings, and a portrait that fails to arrive is not charged.
2. Download the still whose face you picked and attach it in Nano Banana Pro as the photo to edit, set the size to 2K and give one instruction: the same person in the same light, sharper detail. That edit is your hero still.
3. Attach the hero still as the photo to edit and make the six sheet shots, one instruction each.
4. Open a sheet still as the start frame in Kling 3.0 and describe only the motion. Render short first.
5. For a scene the sheet does not show, open Seedance 2.0 Fast, attach up to nine sheet photos and cite them as @Image1, @Image2 in the prompt.

• One photo per edit. The image models here read one reference photo at a time, so a set of angles is not averaged; the best single shot does the work.
• No reference photos on Kling here. Kling 3.0 and Kling O3 Omni take a start frame only on DreamClip; the reference route is Seedance 2.0 Fast, at 480p or 720p.
• Long takes. Past ten seconds from one start frame every model drifts; the fix is shorter clips and cuts, not a longer prompt.
• Two characters in one frame. Each is held worse than one would be; give the shot a clear lead.

Midjourney's reference parameters and Runway's reference tools aim at the same problem inside their own apps, as of October 2026. DreamClip does not run Midjourney or Runway. Here the method is a photo edit on Nano Banana Pro for the stills, and a start frame or Seedance 2.0 Fast's reference photos for the clips, all in one studio.