Updated · 6 min read
Reference images in AI video: keeping a face or a product consistent
The question behind most retakes is consistency: the same person, the same bottle, the same room from shot to shot. Reference images are how you get it. Here is how they behave on each engine on 2036 Studios and the habits that make them work.
How many, and in what order
Seedance 2.0 takes up to 9 image references plus one video reference. Seedance 2.5 takes up to 30 image references and no video. Kling 3.0 takes a first frame and a last frame; Kling 3.0 Omni takes image references as well. Order matters: the first image carries the most weight for identity, so put the hero there. In Marketing Studio formats the first image is always the product and the second is the presenter.
Naming references in the prompt
Prompts can refer to Image 1, Image 2 and Video 1 by position. In the video studio, type @ to insert a tag that resolves to the right phrase when the prompt is sent, so the engine never sees a stray symbol it might paint on screen.
Faces are different
Video engines refuse photoreal faces supplied as plain reference images, as a deepfake safeguard. 2036 Studios handles that two ways. Images you generated on the platform are recognised as platform output and ride as trusted references, sharp. Uploaded photos of real people go through a privacy step before they reach the engine, which keeps the render going but softens detail. If you want a sharp, consistent human character in video, generate them.
Build a character once
Generate three or four images of your character with one locked description: same face, hair, outfit and light, in different angles. Create a character from them. You get an @tag that works in any video prompt on Seedance, and on Kling 3.0 the same character rides as a locked element. The character is reusable across every ad and story you make.
Shape and size rules
- Reference images should be between roughly 2:5 and 5:2 in aspect. Very wide screenshots are cropped for you at upload.
- Reference videos should be 15 seconds or shorter and at least a modest resolution; small clips are upscaled for you.
- Keep one palette and one scale across a pack. A stylised close-up next to realistic wides reads as two worlds.
Continuity between renders
To continue a shot, use the last frame of the previous render as the first frame of the next. The platform can extract it for you. Expect a brief held beat at the join; for music videos and dance, render parallel clips with the same references and cut on a flash instead.
Questions, answered
Why did my reference get blurred?+
It contained a real person's face as an uploaded photo. Generate the character on the platform for a sharp result.
Can I reference a product and a person in one render?+
Yes. Product first, person second, and name them in the prompt as Image 1 and Image 2.
Do references cost extra?+
On Seedance 2.0 standard a reference adds a few credits. Fast, Mini and Seedance 2.5 charge the same either way.