Center stage, under the spotlight, a high-pitched Danjiao (female role in Chinese opera) vocal rises, and the character performs a series of spear movements in rhythm with the drums. This is not a live performance, but a "single-take" creation generated by Seedance 1.5 Pro. While a significant gap remains compared to professional opera, the essence of its sound and form is beginning to take shape.

Today, we officially release the new-generation audio-visual generation model: Seedance 1.5 Pro.

This model supports joint audio-video generation, capable of performing various tasks including text-to-audiovisual synthesis and image-guided audiovisual generation. This means Seedance's video generation is no longer confined to the visual dimension; sound is now naturally integrated.

With Seedance 1.0, we focused on improving the model's "performance floor," optimizing motion generation stability. Seedance 1.5 Pro, while supporting synchronized audio generation, also aims to raise the "performance ceiling" in terms of visual impact and motion effects. Through more ambitious technical approaches, the model achieves breakthroughs in several key areas:

1. Precise Audio-Visual Synchronization & Multilingual/Dialect Support: The model achieves high audio-visual consistency during generation, improving the alignment of character lip movements, intonation, and performance rhythm. It natively supports multiple languages and distinctive dialect accents, capturing their unique speech cadence and emotional expressiveness.
2. Cinematic Camera Control & Dynamic Tension: The model exhibits autonomous camera choreography capabilities, executing complex shots like long tracking shots and Hitchcock zooms. It also achieves professional-grade scene transitions and cinematic toning, significantly enhancing the video's dynamic tension.
3. Enhanced Semantic Understanding & Narrative Coherence: Through improved semantic comprehension, the model achieves better parsing of narrative contexts. It markedly enhances the overall narrative coherence of audio-video segments, providing strong support for professional-grade content creation.

Seedance 1.5 Pro is no longer content with generating simple clips; it treats video and audio as a unified creative whole to meet diverse creative demands. Its grasp of audio-visual synergy, dynamic camera work, and cultural context allows it to demonstrate compelling narrative expression and audio-visual integration across scenarios like film/TV production, short drama generation, advertising, and traditional opera performance.

Next, we'll break down how Seedance 1.5 Pro empowers professional creation through specific scenarios.

1. Subtle and Coherent Narrative Expression for Cinematic Art Creation

Seedance 1.5 Pro shows significant improvement in semantic understanding, enabling it to interpret nuanced human emotions and translate them into expressive artistic output. Coupled with high-precision audio-visual synchronization, the model can deeply integrate speech, visuals, and scene atmosphere, generating consistent and refined presentations that enhance the narrative impact of the content.

In close-up shots, the model demonstrates delicate emotion-capturing ability. Even without dialogue, it can sustain emotional buildup through subtle facial expression changes. For example, in generated cyberpunk-style clips, the model can infer the story background from prompts and meticulously portray character states. The emotional shifts are natural and layered, achieving unity with the environment and musical atmosphere.

Beyond story-driven close-ups, Seedance 1.5 Pro can also organize basic narrative shot sequences based on prompts. For instance, in anime-style creation, the model can generate multiple continuous shots—like fireworks blooming and characters confessing in Japanese—paired with emotional vocal delivery, showcasing smooth narrative logic.

2. Professional Camera Work & Dynamic Tension for Demanding Scenes

Seedance 1.5 Pro is optimized for camera control and dynamic expression, better handling complex, high-motion generation scenarios.

The model can relatively comfortably depict high-dynamic, high-impact action scenes. In a skiing video, through synchronized sound and motion, Seedance 1.5 Pro creates a strong sense of presence: the camera sharply cuts sideways following the skier's trajectory, capturing the spray of snow in detail, realistically conveying the speed and power of extreme sports.

Simultaneously, the model possesses spontaneous camera-scheduling capabilities, able to execute complex camera movements for precision-demanding generation. When simulating a red-carpet premiere scene, it can generate rapid panning shots, creating a sense of hurried excitement and grandeur, paired with clear Chinese female narration to effectively recreate the on-site atmosphere.

In a generated promotional video for a robot vacuum, the scene slowly pushes the camera in a commercial ad style, closely following the product's movement, effectively keeping the focus on it.

3. Multilingual & Dialect Support to Enhance Stylized Performances (e.g., Comedy)

Seedance 1.5 Pro supports multilingual speech generation, producing relatively natural vocal delivery in Chinese, English, Japanese, Korean, Spanish, Indonesian, and more. Particularly within Chinese contexts, it can mimic various dialect accents like Sichuanese and Cantonese, adding more authentic performance texture to short dramas and entertainment content.

For example, a giant panda munching bamboo suddenly "complains" to the camera in a Sichuan accent; the model can match the dialect's cadence and expression, bringing the video to life.

4. Precise Sound Effect Generation to Boost Immersion in Content (e.g., Games)

Beyond human voices, Seedance 1.5 Pro also demonstrates good understanding of ambient sounds and musical atmosphere. The model can overlay environmental audio based on visual content, creating spatial awareness and achieving "what you see is what you hear."

In a pixel-art game clip, the model not only achieves smooth camera movement following a character's run and jump but also synchronously generates fitting 8-bit game sound effects, showcasing audio-visual synergy in fast-paced action.

In a 3D-style game segment, the model generates a richly detailed open world. As the character moves, footsteps and breathing sounds are precisely synchronized, accompanied by the distant low caw of a crow, enhancing the immersive quality of the audio-visual interaction.

Leveraging these capabilities, Seedance 1.5 Pro can effectively support typed creation in film/TV, advertising, short dramas, and anime. Particularly in I2V (Image-to-Video) tasks, the model demonstrates strong style consistency, effectively maintaining character feature stability across multi-shot switching and complex motion, improving the coherence from raw footage to final production.

To objectively evaluate the model's comprehensive abilities, the team established the integrated evaluation benchmark SeedVideoBench 1.5. Co-developed by film directors and technical experts, this benchmark focuses on assessing the model's performance across dimensions like visual complex instruction following, motion stability & vividness, aesthetic quality, as well as audio instruction following, audio-visual synchronization, and audio quality/expressiveness.

Evaluation results show:

1. In video generation, compared to other benchmarked models, Seedance 1.5 Pro demonstrates relatively more accurate understanding of complex instructions for actions and camera work, better matching the narrative and visual style set by prompts. Its dynamic performance is fuller, character close-up expressions are vivid, and complex camera movements are relatively smooth and naturally aligned with the reference image style. The overall visual texture is closer to live-action footage, though motion stability still has room for improvement.
2. In audio generation, Seedance 1.5 Pro performs at an industry-leading level. The model shows stable and balanced performance in audio instruction following, audio-visual sync, and audio quality/expressiveness. It can relatively accurately generate matching human voices and specified sound effects, demonstrating particularly high integrity and pronunciation clarity in Chinese dialogue scenarios, and can respond to various dialect instructions.

Compared to similar models, human voices generated by Seedance 1.5 Pro are relatively more natural with less mechanical feel. Sound effects are more realistic with spatial reverb closer to reality, and audio-visual misalignment is significantly reduced. Although future work will focus on improving its performance in multi-character alternating dialogue and singing scenarios, overall, the model can already be partially applied to narrative scenes driven by Chinese or dialect dialogue, such as short dramas, stage performances.

In terms of technical architecture, Seedance 1.5 Pro adopts a base model design for joint audio-visual generation. Through systematic restructuring of the underlying architecture, data pipeline, post-training, and inference stages, it enhances generalization performance across diverse and complex tasks:

1. Multimodal Joint Architecture: A unified audio-visual generation framework based on an MMDiT architecture enables precise temporal and semantic alignment between visual and auditory streams through deep cross-modal interaction mechanisms.
2. Multi-stage Data Pipeline: A multi-stage data strategy balances audio-visual consistency and motion expressiveness. It significantly enriches video descriptions with professional terminology and incorporates audio descriptions, providing a high-quality, diverse data foundation for high-fidelity audio-visual generation tasks.
3. Refined Post-Training Optimization: High-quality audio-visual datasets are used for Supervised Fine-Tuning (SFT), and RLHF algorithms customized for audio-visual scenarios are introduced. Specifically, multi-dimensional reward models effectively enhance performance in T2V and I2V tasks, comprehensively improving motion quality, visual aesthetics, and audio fidelity.
4. Efficient Inference Acceleration: The multi-stage distillation framework is further optimized, significantly reducing the required Number of Function Evaluations (NFEs). Combined with optimizations like quantization and parallel processing in the inference infrastructure, end-to-end inference speed is accelerated by over 10x while maintaining model performance.

Compared to our previous-generation video generation model Seedance 1.0, Seedance 1.5 Pro marks a crucial step forward in immersive audio-visual experience and production-grade narrative expression.

Relying on its joint architecture and refined training, Seedance 1.5 Pro achieves relatively good adherence to multimodal instructions—showing high potential both in high-dynamic cinematic camera work and in dialect performances requiring precise lip-sync. However, we also note areas for improvement, such as physical stability in highly challenging motions, multi-character dialogue, and singing scenes.

Looking ahead, the Seed team will focus on breaking through longer-duration narrative generation and more real-time on-device experiences, while further strengthening the model's understanding of physical world laws and its multimodal perception capabilities. We hope the Seedance series models will become even more vivid, efficient, and user-aware, empowering creators to break sensory boundaries and realize their audio-visual creativity.