Seedance 2.0 First and Last Frame Control

Seedance 2.0 is ByteDance’s multimodal video generation model for creating and editing videos from text, images, video, and audio references. Atlas Cloud provides text-to-video, image-to-video, and reference-to-video endpoints with synchronized native audio.

Create an Atlas Cloud API key and store it securely as an environment variable. Send an authenticated POST request to the video generation endpoint with the appropriate model ID, prompt, and generation settings. Use the returned prediction ID to retrieve the completed output.

Choose text-to-video when the scene starts from a written prompt. Use image-to-video to animate a first-frame image with an optional last frame, or select reference-to-video when images, videos, or audio must guide identity, motion, style, editing, or extension.

Available inputs depend on the endpoint and include prompts, first and last frames, reference images, reference videos, and reference audio. The full model supports 4 to 15 second outputs, multiple aspect ratios, synchronized audio, bitrate selection, seed control, optional watermarks, and separate last-frame delivery.

Atlas Cloud’s verified standard price is $0.112 per second for each of the three listed family endpoints. This standard rate avoids temporary discount assumptions and gives developers a stable baseline for estimating production costs.

A generation request returns a prediction ID while processing continues asynchronously. Poll the prediction endpoint until the status becomes completed, failed, or timeout, then read the generated media URL from the outputs array. Avoid submitting duplicate jobs while the original request is still processing.

Set the resolution field to 4k when calling a full Seedance 2.0 endpoint, since Fast and Mini variants do not support this option. Native 4K output is 3840 by 2160 at 16:9 and uses 10-bit H.265 or HEVC encoding. If playback fails in a browser, open the file in an HEVC-compatible player or transcode it for broader compatibility.

Direct reference image or video uploads containing real human faces are not supported through the ordinary reference workflow. Where available, use platform-approved virtual portraits or properly authorized portrait assets, and obtain the subject’s consent before generating likeness-based content.

Start by checking the model ID, required prompt, duration, resolution, aspect ratio, and accessibility of every referenced media URL. Real-face restrictions and content safety checks can also reject an otherwise valid request. If submission succeeds, inspect the prediction status and error response before retrying.