Seedance 1.5

About Seedance 1.5 Pro

Launched by ByteDance in December 2025, Seedance 1.5 Pro is a fundamentally different approach to AI video generation: instead of generating silent video and adding audio as a separate step, it generates both streams simultaneously. The model's Dual-Branch Diffusion Transformer uses 4.5 billion parameters split into two parallel branches—one for video frames, one for audio waveforms—connected by a cross-modal joint module that ensures synchronization at the millisecond level. The result is video where a character's lips match spoken words, footsteps click exactly when feet hit the ground, and ambient soundscapes match the visual density of the scene—all from a single prompt.

Native Audio-Visual Synchronization

The audio system achieves phoneme-level lip sync across eight languages: English, Mandarin, Japanese, Korean, Spanish, Portuguese, Indonesian, and regional Chinese dialects including Cantonese and Sichuanese. This is not post-hoc dubbing. The model reads emotional context ("she whispers nervously" vs. "she speaks confidently") and adjusts vocal delivery, facial expressions, and body language together. Multi-speaker conversations with distinct vocal identities are supported, though two-person exchanges produce the most stable results. Audio generation is optional via API flag, and failed generations automatically refund credits.

Generation Capabilities and Specs

The model was trained on approximately 100 million minutes of audio-video clips using a multi-stage pipeline with automated filtering, caption generation, and curriculum learning, followed by Supervised Fine-Tuning and Reinforcement Learning from Human Feedback (RLHF).

Known Limitations

Seedance 1.5 Pro has specific weaknesses to account for in production workflows. High-speed motion and fast camera pans often produce artifacts or instability. Hand close-ups are a persistent challenge—fingers may appear in anatomically incorrect positions, which affects product demonstrations requiring precise manipulation. Singing performances are unreliable; the model handles spoken dialogue well but musical lip sync lacks the timing precision needed for believable performances. Three or more characters speaking simultaneously creates sync degradation—two-speaker dialogues are the practical limit. Finally, the 12-second maximum means longer content must be stitched from multiple generations, which can introduce subtle visual discontinuities between clips.

Tips for Best Results

• Use film terminology for camera work: "dolly push-in" is more reliable than "moves closer"; "crane shot rising" beats "goes up". The model trained on cinematography vocabulary and responds precisely to it.
• Integrate audio cues in the main prompt: You do not need a separate audio description field. Write "her heels clicking on marble floors" within the scene description and the model generates matched sound.
• Use draft mode for iteration: Testing ten prompt variations at draft quality costs significantly less than two full-resolution renders. Draft mode mirrors traditional pre-visualization workflows.
• Provide a reference image for character consistency: When generating multiple clips with the same character, supply a reference image. The model uses it as an anchor to preserve facial features, clothing, and body proportions across generations.
• Match aspect ratio to platform at generation time: Create separate renders for 16:9 (YouTube), 9:16 (TikTok/Reels), and 1:1 (Instagram feed) rather than cropping after the fact—the model composes each ratio natively.