How to Keep a Character Consistent Across AI Video Shots
MyDreamVid · 2026-09-11
Generate the same character twice from text alone and you'll get two strangers wearing similar clothes. That's the core frustration of AI filmmaking — and it has a reliable fix: stop describing your character in words and start showing the model who they are.
The tool: reference images
Reference-to-video mode lets you attach images alongside your prompt and refer to them by handle: "@Image1 is the main character; she walks along the harbour wall at dawn." The model then keeps the face, hair, outfit and overall design of @Image1 instead of re-inventing them. Seedance 2.5 accepts up to 30 reference images in a single generation; the 2.0 family accepts up to 9.
The workflow, step by step
- Create the character once. Generate a clean portrait in the image workspace (Seedream 5.0 Pro) — neutral background, clear view of face and outfit. Iterate until it's exactly your character; this image is now the single source of truth.
- Do the same for each important location — one wide establishing shot per setting.
- Reference them in every shot. In each video prompt, attach the portrait and mention it: "@Image1 stands at the window, rain outside." The wording matters less than the reference itself.
- Add a style note to every shot. A consistent line like "hand-painted anime look, warm palette, soft film grain" keeps the rendering style uniform even when the content changes.
In the film workspace steps 1–4 are built in: characters and settings live in a cast list, portraits attach to cast members, and any shot can reference them — no re-uploading per shot.
Chaining shots with the last frame
For continuous action — a character walking through a door, a camera following into the next room — reference images aren't enough, because each shot still starts from scratch. The fix is last-frame chaining: take the final frame of shot N and use it as the first frame of shot N+1 (image-to-video). The film workspace does this automatically when you mark a shot as "continue from previous". Motion reads as unbroken because, visually, it is.
Motion and rhythm references
References aren't limited to images. A short reference video clip can drive camera movement or choreography ("@Video1 drives the camera motion"), and a reference audio file can set rhythm. Two practical notes: reference clips must be at least 2 seconds long, and reference video/audio is billed on top of the output by its own length and resolution — the quote shows the exact addition before you generate.
One honest limitation
Reference images of real, recognizable human faces are declined — every Seedance model refuses to animate real people, and that includes reference photos. Consistency workflows work with invented characters: illustrations, anime, stylised 3D, product mascots.
Frequently Asked Questions
Do reference images cost extra?
In video generation, reference images add nothing to the billed seconds (reference video and audio do add cost, by their own length). In image-to-image generation, the first reference image is free and each extra one adds 1 credit per generated image.
My character still drifts between shots. What else can I do?
Use the same portrait in every shot, keep the style note identical, describe the outfit in words as well ("red scarf, brown coat") as a belt-and-braces measure, and prefer Seedance 2.5 — it follows references most faithfully. For continuous sequences, chain shots by last frame instead of generating them independently.
Can I use several characters in one shot?
Yes — attach several portraits and refer to them separately: @Image1, @Image2. Keep the prompt explicit about who does what.