Seedance 1.5 Pro

Seedance 1.5 Pro is built on a dual-branch Diffusion Transformer. One branch generates the video frames while the other generates the audio, and a cross-modal joint module keeps the two aligned at every step of the diffusion process. That is why speech, lip movement and sound effects come out of a single pass instead of being generated silent and dubbed afterwards, which is how most earlier video models worked.

ByteDance trained it with a multi-stage pipeline that prioritizes audio-video coherence and motion expressiveness, then refined it with supervised fine-tuning on curated audio-video datasets and reinforcement learning from human feedback using a multi-dimensional reward model. A separate acceleration framework, built on multi-stage distillation and infrastructure work, speeds up inference by more than 10x compared with the unaccelerated model.

Compared with Seedance 1.0, which was a silent video model focused on making motion stable and reliable, 1.5 Pro pushes the ceiling instead of the floor. It adds native audio with lip sync across English, Mandarin, Japanese, Korean, Spanish and Indonesian, plus Chinese dialects like Cantonese and Sichuanese, and it handles more ambitious cinematography such as dolly zooms, long takes and multi-shot narrative sequences from a single prompt.

ByteDance is open about the current limits: the release notes acknowledge room for improvement in physical stability during complex motion, multi-character dialogue and singing performances. The model shipped in December 2025 and is available through Volcano Engine as well as third-party API platforms such as fal and Replicate.