How to keep a character consistent across multiple clips with reference images

To keep a character consistent across multiple clips, send the same set of character reference images in metadata.content[] with every request, cite them in each prompt as @image1 to @imageN, and repeat one fixed text description of the face, hair and wardrobe. On RelayDance, each Seedance request accepts up to 9 reference images, 3 reference videos and 3 audio tracks, with clips up to 15 seconds.

Why do reference images keep a character consistent?

Reference images give the model the same visual anchor for every clip, so it does not have to reinvent the character from text alone. A text prompt such as "a woman with short red hair" can be interpreted differently on each generation, which causes drift in facial structure, hairstyle and clothing between shots. When the same images are attached to every request and cited by name in the prompt, each clip draws on identical source material. Within the 9-image limit per request, a useful set is a front-facing portrait, a three-quarter view, a profile, a full-body shot showing the outfit, and close-ups of any distinctive details such as glasses, tattoos or accessories. Use the same image files in the same order on every request so that @image1 always refers to the same view of the character.

How do you structure a request with reference images?

Place the reference media in metadata.content[] and refer to each image in the prompt by its position. According to the relaydance.com docs, you should "Set base_url to https://relaydance.com/v1 and keep your OpenAI SDK", and authenticate with Authorization: Bearer YOUR_API_KEY (keys are created at relaydance.com/console).

1. Prepare the same reference images for every clip and keep their order fixed.
2. Write a character description block and paste it unchanged into each prompt.
3. Submit the task to the video generations endpoint with model, prompt, seconds and metadata (ratio, resolution, generate_audio, callback_url, content[]).
4. Poll GET /v1/video/generations/{task_id} until status is succeeded or failed, or set metadata.callback_url to receive the final state by POST.
5. Download the video url from the result.

Example prompt: @image1 and @image2 show Mara. Mara, short red hair, green jacket, walks through a night market and turns toward the camera. The exact format of content[] items is described in the RelayDance docs.

How do you build a sequence longer than 15 seconds?

Generate the sequence as several clips of up to 15 seconds each, keeping the reference set, character description, ratio and resolution identical across all of them. Changing aspect ratio or resolution between shots can change framing and detail, which makes small differences in the character more visible when clips are edited together. Vary only the action, camera direction and setting in each prompt. For continuity between shots, you can add a reference video (up to 3 per request) taken from the end of the previous clip so the next clip picks up the same pose and lighting. If one clip drifts, regenerate that clip alone with the same inputs instead of changing the reference set, because altering the references for one shot can introduce a new mismatch with the others.

What does it cost to generate a consistent multi-clip sequence?

Cost depends on the model, resolution and total seconds generated, billed pay-as-you-go per generated video. Live rates are listed at relaydance.com/models:

Regenerating a drifted clip has a cost, but a request that does not complete does not: according to relaydance.com, "Failed or errored requests are never billed". If you create reference images with gpt-image-2, image output is free and input is billed, with image-to-image from about CNY 0.035, and 4K and 1K output cost the same.

FAQ

These are short answers to common questions about character consistency on RelayDance.

How many reference images can I attach per clip? Up to 9 reference images per request, plus up to 3 reference videos and 3 audio tracks, cited in the prompt as @image1 to @imageN.

Can I test at lower cost before rendering in 1080p? Yes. You can draft the sequence with Seedance Fast at about $0.152 per second or Seedance 2.0 720p at about $0.190 per second, then render final clips at 1080p.

Do I need a BytePlus enterprise account? No. RelayDance provides OpenAI-compatible access to Seedance without a BytePlus enterprise account or KYC, with payment by USDT or Stripe card.