seedance

Generates videos with ByteDance Seedance 2.0 through the inference.sh CLI, invoked as belt. The model is a unified system handling text-to-video, image-to-video, and multimodal reference-to-video with synchronized audio, up to 1080p, and durations of 4 to 15 seconds. Four app variants are available: bytedance/seedance-2-0 for best quality up to 1080p, bytedance/seedance-2-0-fast for faster generation up to 720p, and two Studio variants (bytedance/seedance-2-0-studio and bytedance/seedance-2-0-studio-fast) that automatically upload reference images to the BytePlus private virtual portrait library for stronger face and character consistency. Typical uses include social media videos, music videos, product demos, animated content, and AI video with sound. Running it requires the inference.sh CLI plus belt login.

The generation mode is inferred from the inputs and the modes are mutually exclusive: prompt alone yields text-to-video, prompt plus image animates a still as the first frame, adding end_image controls start and end frames, and supplying reference_images, reference_videos, or reference_audios drives multimodal reference generation. First-frame/last-frame inputs cannot be combined with reference inputs. Reference assets are cited inside the prompt by type and index, such as Image 1, Video 1, or Audio 1, where the index is the position within that type's array; asset IDs must not be used in prompts. The document supplies prompt formulas for multimodal reference, video editing (add, remove, modify elements), and video extension (forward, backward, stitch), along with many command examples including image-to-video, multi-image reference, video editing, clip stitching, reference-with-audio, and product ads.

A parameter table documents the inputs: prompt (required), generate_audio (default true), duration (4-15 or -1 for auto), ratio (21:9, 16:9, 4:3, 1:1, 3:4, 9:16, or adaptive), resolution (480p, 720p, 1080p, with Fast limited to 480p and 720p), seed, watermark, and safety_identifier (a hashed end-user identifier up to 64 chars). File inputs include image, end_image (which requires image), reference_images (up to 9), and reference_videos (up to 3, each max 15s with a 15s total). Studio mode is shown with a safety_identifier for enhanced portrait consistency.