Faces look warped in image to video results: input photo tips that keep the person recognizable

Warped faces in image to video results usually trace back to the reference photo, not the model. To keep a person recognizable, supply a clear front facing photo, place it in metadata.content[], and cite it in the prompt as @image1. RelayDance accepts up to 9 reference images per request. Because failed or errored requests are never billed, you can retest inputs at no cost on Seedance 2.0 720p from about $0.190 per second.

Why input photos cause warped faces

Warped faces most often come from low resolution, off angle, or partially occluded reference photos. When the face occupies a small fraction of the frame or is turned sharply, the model has fewer stable features to preserve across motion, so identity drifts. Reference media are supplied in metadata.content[] and cited in the prompt as @image1 to @imageN, with a limit of 9 reference images, 3 reference videos, and 3 audio tracks per request. Clips can run up to 15 seconds. Longer clips give the face more frames to shift, so keeping the subject centered and well lit reduces the chance of distortion over the full duration.

Input photo checklist that keeps identity stable

Use a sharp, evenly lit, front facing photo where the face fills a large portion of the frame.

1. Choose a front facing shot with both eyes visible and neutral expression.
2. Ensure even lighting with no strong shadows across the face.
3. Keep the face large in the frame and avoid heavy background clutter.
4. Remove obstructions such as sunglasses, hair over the eyes, or hands near the face.
5. Place the file in metadata.content[] and reference it as @image1 in the prompt.
6. Keep the requested seconds short during testing, then extend up to the 15 second limit.

Because retries are free, iterate on the photo before spending on longer outputs. See live rates at relaydance.com/models.

Test cheaply, then scale resolution

Start on a lower cost tier to validate face stability, then move to higher resolutions once the input works.

Run identity checks on Seedance Fast at about $0.152 per second, then re-render on 1080p at about $0.470 per second when the face holds. According to the official relaydance.com docs, "Failed or errored requests are never billed", so failed attempts during tuning add no cost.

Submit and poll with the OpenAI SDK

You can integrate with existing OpenAI tooling without a new client library.

According to the official relaydance.com/docs, "Set base_url to https://relaydance.com/v1 and keep your OpenAI SDK". Submit a task with POST /v1/video/generations, passing model, prompt, seconds, and metadata (including ratio, resolution, and your content[] reference media). Poll GET /v1/video/generations/{task_id} until status is succeeded or failed, then read the video url from the result. For asynchronous delivery, set metadata.callback_url and the final state is POSTed to your server. Authenticate with Authorization: Bearer YOUR_API_KEY. Full details are at relaydance.com/docs.

FAQ

How many reference images can I attach to one request?

Up to 9 reference images, plus 3 reference videos and 3 audio tracks per request. Place them in metadata.content[] and cite each as @image1 through @imageN in the prompt.

Do I pay for a render that fails partway through?

No. According to the official relaydance.com docs, "Failed or errored requests are never billed", so tuning your input photo does not add cost when a request errors.

What is the maximum clip length?

Clips can run up to 15 seconds. For face stability testing, keep seconds short first, then extend once the reference photo produces a recognizable result.