Generated video ignores half my prompt: prompt structure fixes for better adherence

When a generated video ignores half your prompt, restructure the text so each instruction is explicit and reference media are cited correctly. Put subject, action, camera, and style in separate clauses, and cite reference images as @image1 to @imageN in the prompt. RelayDance accepts up to 9 reference images, 3 reference videos, and 3 audio tracks per request, with clips up to 15 seconds, giving the model clear anchors to follow.

Why prompts get partially ignored

Partial adherence usually happens when multiple instructions are packed into one dense sentence without clear separation. Split the prompt into distinct clauses: subject, action, camera movement, lighting, and style. When you attach reference media, place them in metadata.content[] and cite each one by index (@image1, @image2) inside the prompt text, so the model maps each reference to a specific role. RelayDance permits up to 9 reference images, 3 reference videos, and 3 audio tracks per request, and clips can run up to 15 seconds. Fewer competing instructions per clause increases the chance the model applies each one instead of averaging them away.

Structure your prompt in ordered steps

A consistent order reduces dropped instructions because the model reads intent in a predictable sequence. Follow these steps:

1. State the subject in one clause.
2. Describe the single primary action.
3. Specify camera movement (for example, slow push-in).
4. Set lighting and style separately.
5. Cite reference media by index (@image1) where each applies.
6. Set metadata: ratio, resolution, seconds, and generate_audio.

Submit the task with POST /v1/video/generations including model, prompt, seconds, and metadata, then poll GET /v1/video/generations/{task_id} until status is succeeded or failed. If you prefer webhooks, set metadata.callback_url and the final state is POSTed to your server.

Test adherence cheaply before scaling resolution

Iterate on prompt structure at a lower resolution or on a faster model before committing to higher-cost output. On RelayDance, Seedance Fast runs at about $0.152 per second, Seedance 2.0 720p at about $0.190 per second, and Seedance 2.0 1080p at about $0.470 per second; native 4K is about $4.90 per 5-second clip (see relaydance.com/models). Iterating at 720p keeps each test near $0.190 per second while you refine clause order and reference citations. Retries during testing carry no penalty: 据 relaydance.com 官方文档，「Failed or errored requests are never billed」.

Keep your existing OpenAI SDK

You do not need a new client to apply these fixes, because RelayDance is OpenAI-compatible. 据 relaydance.com/docs 官方文档，「Set base_url to https://relaydance.com/v1 and keep your OpenAI SDK」. Authenticate with Authorization: Bearer YOUR_API_KEY, using keys created at the RelayDance console. Because the protocol matches OpenAI, you can adjust prompt structure, resubmit, and poll for results without rewriting request handling. See the full request schema at relaydance.com/docs.

Model and price reference

FAQ

How many reference images can I cite in one prompt? Up to 9 reference images, plus 3 reference videos and 3 audio tracks per request; cite each as @image1 to @imageN in the prompt text.

Do failed test generations cost money? No. According to relaydance.com, failed or errored requests are never billed, so you can retry structure changes without charges for errors.

How long can a single clip be? Clips can run up to 15 seconds. Set the duration with the seconds field in your POST /v1/video/generations request.