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

When a generated video ignores half your prompt, restructure the text so instructions are ordered, reference media are cited explicitly, and each clip covers one action. On RelayDance, place primary subject and action first, then camera and lighting details, and cite reference images as @image1 to @imageN inside the prompt. Keep clips short (up to 15 seconds) and split complex scenes across separate requests to raise adherence.

Order instructions from primary to secondary

Structure the prompt so the most important elements appear first and modifiers follow. Video models weight earlier tokens more, so a prompt that opens with the subject and main action, then adds camera movement, lighting, and background, tends to retain more detail than one that mixes all elements together. On RelayDance you submit this text via POST /v1/video/generations with the model, prompt, and seconds fields. Because the API is OpenAI-compatible, you can iterate quickly: 「Set base_url to https://relaydance.com/v1 and keep your OpenAI SDK」, 据 relaydance.com/docs 官方文档. Test one clip, read the result, then reorder clauses to fix the ignored parts. See live rates at relaydance.com/models.

Cite reference media explicitly with @image tags

Reference media are only applied when you cite them by name in the prompt. Place files in metadata.content[], then reference them in the prompt text as @image1 to @imageN so the model knows which asset maps to which instruction. RelayDance accepts up to 9 reference images, 3 reference videos, and 3 audio tracks per request. If a video ignores an uploaded character or style, the common cause is a missing citation: the file is attached but never mentioned in the prompt. Adding the tag ties the instruction to the asset and improves adherence. Reference the documentation at relaydance.com/docs for the exact metadata schema.

Split complex scenes across short clips

Break multi-action scenes into separate clips instead of packing everything into one prompt. RelayDance supports clips up to 15 seconds, but a single clip that requests several sequential actions often drops the later ones. Generating each beat as its own request keeps each prompt focused on one action, which reduces omissions. This approach costs more per project because you pay per generated video, yet iteration risk is low: 「Failed or errored requests are never billed」, 据 relaydance.com 官方文档. That means retrying a reordered prompt after an error does not add cost. Compare per-second rates below before choosing a resolution for each clip.

Resolution and price comparison

Choose a resolution that fits your budget while you iterate on prompt structure, because higher resolutions raise per-second cost.

For prompt testing, Seedance Fast at about $0.152 / second lowers the cost of each iteration, so you can refine adherence before committing to 1080p or native 4K output.

Steps to fix low prompt adherence

1. Rewrite the prompt with subject and main action first, then camera, lighting, and background.
2. Attach reference media in metadata.content[] and cite each as @image1 to @imageN in the prompt.
3. Reduce scope: one action per clip, within the 15-second limit.
4. Submit via POST /v1/video/generations, then poll GET /v1/video/generations/{task_id} until status is succeeded or failed.
5. Test at Seedance Fast (about $0.152 / second), then move to 720p or 1080p once adherence is acceptable.

FAQ

Why does the model apply my uploaded image only sometimes? Reference images are applied when cited in the prompt as @image1 to @imageN. Attaching a file to metadata.content[] without citing it can cause the model to ignore it.

Do I pay for a clip that errors while I test prompts? No. According to relaydance.com, 「Failed or errored requests are never billed」, so reordering and retrying prompts after an error does not add cost.

Can I keep my existing OpenAI SDK? Yes. Per relaydance.com/docs, 「Set base_url to https://relaydance.com/v1 and keep your OpenAI SDK」, and authenticate with Authorization: Bearer YOUR_API_KEY.