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

Warped faces in image to video output usually trace back to the input photo, not the model. To keep a person recognizable on RelayDance, supply a sharp, well lit, front facing photo where the face fills a large share of the frame, avoid heavy occlusion (sunglasses, hats, hands), and place the reference in metadata.content[], citing it in the prompt as @image1. RelayDance is OpenAI-compatible; failed requests are never billed, so iterating on inputs carries no cost for errors.

Why input photos cause warped faces

Warping starts when the reference photo gives the model too little stable facial detail to preserve across frames. Low resolution, motion blur, extreme angles, and partial occlusion force the model to interpolate, and interpolation distorts identity. Use a single subject, front facing or near front facing, with even lighting and no strong shadows across the face. Keep the face large in the frame so features are clearly defined. On RelayDance you attach references through metadata.content[] and reference them in the prompt with tags @image1 to @imageN, supporting up to 9 reference images, 3 reference videos and 3 audio tracks per request, with clips up to 15 seconds.

Steps to prepare a reference photo

Follow a fixed checklist before every image to video submission to reduce distortion.

1. Choose a sharp, in focus photo with the face occupying a large area of the frame.
2. Prefer front facing or near front facing angles; avoid profile and extreme tilts.
3. Remove occlusions: sunglasses, hats, hair over the eyes, and hands on the face.
4. Use even, diffuse lighting; avoid harsh side light and blown highlights.
5. Place the file in metadata.content[] and cite it in the prompt as @image1.
6. Submit with POST /v1/video/generations, then poll GET /v1/video/generations/{task_id} until status is succeeded or failed.

How iteration stays low cost while you test inputs

Testing several input photos is inexpensive because you only pay for successful generations. According to the official relaydance.com docs, 「Failed or errored requests are never billed」, so a rejected or errored attempt does not add to your bill. Billing is pay-as-you-go per generated video, with live per-model rates at relaydance.com/models. For reference, Seedance 2.0 720p is about $0.190 per second and 1080p is about $0.470 per second, Seedance Fast is about $0.152 per second, and Seedance native 4K is about $4.90 per 5-second clip. Higher resolution preserves finer facial detail, which can reduce visible warping.

Resolution and model rate comparison

Higher resolution options carry higher per second rates but retain more facial detail.

If a person still looks warped at 720p, moving to 1080p or 4K gives the model more pixels to anchor identity. Note that for gpt-image-2, image output is free and only input is billed, with 4K and 1K output costing the same.

Connecting and submitting requests

RelayDance uses an OpenAI-compatible protocol, so existing tooling works with a base URL change. According to the official relaydance.com/docs, 「Set base_url to https://relaydance.com/v1 and keep your OpenAI SDK」. Authenticate with Authorization: Bearer YOUR_API_KEY, using keys created in the console. Submit an image to video task with POST /v1/video/generations, including model, prompt, seconds, and metadata (ratio, resolution, generate_audio, callback_url, and content[]). To receive results without polling, set metadata.callback_url and the final state is POSTed to your server. See relaydance.com/docs for full request fields.

FAQ

Does a higher resolution always fix warped faces? Not always. Resolution helps by preserving detail, but a blurry or occluded input photo will still distort. Fix the input first, then raise resolution if needed.

Am I charged if a generation fails while I test photos? No. Per relaydance.com, failed or errored requests are never billed, so iterating on inputs carries no error cost.

How many reference images can I attach? Up to 9 reference images, 3 reference videos and 3 audio tracks per request, cited in the prompt as @image1 to @imageN, with clips up to 15 seconds.