ByteDance Seedance family guide: 1.0 motion stability, 1.5 Pro native audio, and 2.0 multimodal reference control. Updated July 2026.
Seedance: ByteDance AI Video Generation Family Guide
Developer: ByteDance Seed Lab · Family launched: June 2025 · Current versions: 1.0, 1.5 Pro, 2.0 · Peak leaderboard: #1 Text-to-Video and Image-to-Video (multiple periods)
What Is Seedance?
Seedance is ByteDance's family of AI video generation models, developed by the company's Seed Lab and launched publicly in June 2025. The family represents ByteDance's approach to production-scale video generation: models designed not just to produce impressive results in ideal conditions, but to deliver consistent, controllable output across the varied, messy, real-world prompts that production workflows actually require.
The name itself reflects the family's positioning. "Seed" references ByteDance's Seed Lab, the research division responsible for the family's development. "Dance" suggests motion, rhythm, and the choreographed precision the models are designed to deliver. Together, the name positions the family as ByteDance's systematic answer to video generation, not a one-off research demo.
From the first public release in June 2025 through mid-2026, Seedance models have held the #1 position on the Artificial Analysis Video Arena leaderboard across text-to-video and image-to-video categories for multiple extended periods, competing directly with Google's Veo, Kuaishou's Kling, and OpenAI's Sora (discontinued by OpenAI, announced March 24 and shut down April 26, 2026). The family's sustained competitive positioning established ByteDance as one of the three primary players in production-scale AI video generation.
What makes Seedance notable is not just the leaderboard performance but the architectural progression. Each version in the family solves a different production constraint:
- Seedance 1.0 (June 2025) established motion stability and inference speed as the baseline
- Seedance 1.5 Pro (December 2025) introduced native audio-visual joint generation
- Seedance 2.0 (February 2026) unified multimodal reference control and platform-level super-resolution to 1080p
This is a family designed to address the full production stack, not just the most photogenic use case.
The Evolution: From Motion Stability to Multimodal Control
Seedance 1.0: The Performance Floor (June 2025)
Seedance 1.0 launched on June 11, 2025, and immediately took the #1 position on both the Artificial Analysis text-to-video and image-to-video leaderboards, beating Google's Veo 3 and Kuaishou's Kling 2.0 by more than 100 Elo points on the image-to-video task.
The defining achievement was not the leaderboard placement but what ByteDance called raising the "performance floor": making it harder to get a bad result from a reasonable prompt. Earlier diffusion-based video models could produce stunning outputs under ideal conditions, but they also produced jittery motion, prompt drift, and morphing subjects with alarming frequency. Seedance 1.0 reduced those failure modes systematically.
Core capabilities:
- Native 1080p generation at 24 FPS
- Text-to-video and image-to-video modes
- Native multi-shot generation via natural language shot labeling
- 10× inference speedup through distillation (5-second clip in 41.4 seconds on NVIDIA L20)
- Strong spatiotemporal coherence and motion plausibility
- Silent generation (no native audio)
Technical foundation: Latent Diffusion Transformer with multi-source data curation, interleaved multimodal positional encoding, and post-training optimization through fine-grained supervised fine-tuning and video-specific RLHF.
The "performance floor" achievement: Subjects stay coherent across frames. Camera movement behaves predictably. Physical relationships hold. Secondary motion follows primary motion with appropriate delay. These are not edge cases but the baseline.
Technical paper: arXiv:2506.09113
Seedance 1.5 Pro: The Performance Ceiling (December 2025)
Seedance 1.5 Pro launched on December 16, 2025, as ByteDance's first audio-visual joint generation model. Where 1.0 focused on raising the floor, 1.5 Pro aimed to raise the "performance ceiling": elevating the visual impact, audio-visual synchronization, and narrative coherence beyond what stable motion alone could deliver.
The architectural shift was fundamental. Seedance 1.5 Pro uses a dual-branch Diffusion Transformer that generates audio and video simultaneously in a shared latent space through a cross-modal joint module. Audio is not added after video generation but is part of the same denoising process. This ensures that lip sync, foley, and ambient sound are causally linked to visual motion during generation, not matched afterward.
Core capabilities:
- Native audio-visual joint generation (dialogue, SFX, ambient, BGM)
- Phoneme-level lip sync across 8+ languages and dialects (Mandarin, English, Japanese, Korean, Spanish, Indonesian, Sichuanese, Cantonese)
- Dual-branch architecture with cross-modal joint module
- Reference audio-to-video mode (visual pacing driven by provided audio track)
- 1080p output at 24 FPS
- Cinematic camera control with dynamic movement responding to narrative pacing
The "performance ceiling" achievement: When a character speaks, their mouth moves in sync with phonemes, not approximate syllables. Foley responds to visual motion because both are generated together. Music and visual pacing align naturally. Multilingual and dialect support removes the need for separate dubbing pipelines.
Technical paper: arXiv:2512.13507
Seedance 2.0: Unified Multimodal Control (February 2026)
Seedance 2.0 launched on February 12, 2026, as the first Seedance model built on a unified multimodal architecture. The shift was not incremental but a fundamental rethinking of how the model accepts and processes input.
Seedance 2.0 processes text, images, audio, and video as inputs simultaneously, reasoning over all of them together during generation. A single generation request can include up to 9 reference images, 3 video clips, and 3 audio files alongside a text prompt. These inputs are not processed sequentially as independent conditioning signals but are integrated into a unified creative decision before the first frame is generated.
Core capabilities:
- Unified multimodal architecture (text, image, audio, video inputs processed jointly)
- Up to 9 reference images + 3 video clips + 3 audio files per generation
- Native audio-visual joint generation with phoneme-level lip sync (8 languages)
- Native resolutions 480p and 720p with platform super-resolution to 1080p
- Multi-shot sequential generation with natural language shot labeling
- First/last frame control in image-to-video mode
- Two variants: Standard (quality-optimized) and Fast (speed-optimized)
The multimodal control achievement: Brand teams can supply product photography, a reference video capturing the visual tone they want, and a brief describing the scene. Seedance 2.0 generates output that honors all three simultaneously: character identity from the product images, visual language from the reference clip, and action from the text, processed as a unified creative problem.
Technical paper: arXiv:2604.14148
Version Comparison: Which Model for Which Workflow
| Feature | Seedance 1.0 | Seedance 1.5 Pro | Seedance 2.0 |
|---|---|---|---|
| Released | June 11, 2025 | December 16, 2025 | February 12, 2026 |
| Core focus | Motion stability baseline | Audio-visual joint generation | Unified multimodal reference control |
| Native resolution | 1080p | 1080p | 480p and 720p |
| Platform output | N/A | N/A | Up to 1080p (via super-resolution) |
| Clip duration | 2 to 12 seconds | 2 to 12 seconds | 4 to 15 seconds |
| Audio output | ❌ Silent | ✅ Native (dialogue, SFX, ambient, BGM) | ✅ Native (dialogue, SFX, ambient, BGM) |
| Lip sync | ❌ No | ✅ Phoneme-level (8+ languages) | ✅ Phoneme-level (8 languages) |
| Reference inputs | Text + single image | Text + single image + audio reference | Text + 9 images + 3 videos + 3 audio files |
| Multi-shot | ✅ Natural language shot labeling | ✅ Natural language shot labeling | ✅ Natural language shot labeling |
| Inference speed | 10× speedup (fastest) | 10× speedup | Standard (quality) or Fast (speed) |
| Best for | Fast iteration, silent B-roll, I2V product animation | Multilingual dialogue, audio-synced content, narrative films | Reference-heavy brand work, multimodal input control |
Shared Architecture Foundations
While each version introduces distinct capabilities, all Seedance models share core architectural principles:
Latent diffusion backbone. All versions process video in latent space using transformer-based architectures trained on curated multi-source datasets with precision video captioning.
Multi-stage training pipeline. Pre-training on large-scale datasets, followed by supervised fine-tuning on high-quality subsets, followed by reinforcement learning from human feedback (RLHF) with multi-dimensional reward models.
Native multi-shot generation. All versions support natural language shot labeling: "Shot 1: [description]. Shot 2: [description]. Shot 3: [description]." generates a sequence with consistent character identity and lighting across cuts.
Inference acceleration through distillation. All versions achieve approximately 10× inference speedup compared to baseline diffusion approaches through multi-stage distillation and system-level optimizations.
Text-to-video and image-to-video modes. All versions support both T2V (from text prompts) and I2V (from source image as first frame anchor).
These shared foundations ensure that the core motion stability, prompt adherence, and structural coherence defined by Seedance 1.0 remain consistent across the family, even as newer versions add audio, multimodal reference control, and higher resolution.
Production Use Cases by Version
Fast Iteration and Concept Testing → Seedance 1.0
When audio isn't part of the early workflow and speed matters, Seedance 1.0 remains the practical choice for high-volume prompt testing and creative direction exploration. The inference speed (5-second clip in 41.4 seconds) and lower credit cost make it viable for iterating through multiple visual directions before committing to a final approach.
Best for: B-roll generation, product animation from stills, storyboard previsualization, rapid concept testing where audio will be added in post.
Multilingual and Dialogue-Driven Content → Seedance 1.5 Pro
When the project requires native audio and video generated together, particularly for multilingual localized content or dialogue-driven narrative, Seedance 1.5 Pro's phoneme-level lip sync across 8+ languages and dual-branch audio-visual architecture make it the direct path from script to synchronized output.
Best for: Multilingual advertising, character-driven short films, audio-synced social content, music video generation with reference audio-to-video mode.
Reference-Heavy Brand Work and Multimodal Input Control → Seedance 2.0
When the project requires consistent character identity across shots, specific camera language extracted from a reference reel, audio tone matching a reference track, visual style anchoring to existing brand imagery, or the richest multimodal reference surface available, Seedance 2.0's unified multimodal input architecture makes it the appropriate model.
Best for: Brand and product content with multiple reference assets, short-form narrative requiring tight reference control, pre-production previsualization with rich reference inputs, commercial work requiring multimodal reference consistency.
Leaderboard Performance and Competitive Positioning
Seedance models have held the #1 position on the Artificial Analysis Video Arena across multiple categories and time periods:
Seedance 1.0 (June 2025):
- #1 Text-to-Video at launch
- #1 Image-to-Video at launch (100+ Elo point lead over second place)
Seedance 2.0 (February to April 2026):
- #1 Text-to-Video (with audio): Elo 1,216
- #1 Image-to-Video (with audio): Elo 1,193
- #1 Image-to-Video (no audio): Elo 1,344
- #2 Text-to-Video (no audio): Elo 1,274 (behind Happy Horse 1.0 at 1,333 to 1,357)
As of mid-2026, Seedance 2.0 holds the top audio-inclusive positions and the top image-to-video position across audio and no-audio categories. Happy Horse 1.0 leads in text-to-video no-audio by approximately 60 Elo points, but trails Seedance 2.0 in the with-audio category by 14 points, reflecting Seedance 2.0's stronger native audio quality.
These rankings come from blind pairwise user comparisons on the Artificial Analysis Video Arena, where users choose between unlabeled videos without knowing which model produced which clip. The sustained competitive positioning across multiple model releases establishes Seedance as one of the three primary families in production-scale video generation, alongside Veo and Kling.
Technical Papers and Official Sources
Each Seedance version is documented in a formal technical paper filed on arXiv:
- Seedance 1.0: arXiv:2506.09113, "Exploring the Boundaries of Video Generation Models"
- Seedance 1.5 Pro: arXiv:2512.13507 — "Seedance 1.5 pro: A Native Audio-Visual Joint Generation Foundation Model"
- Seedance 2.0: arXiv:2604.14148, "Advancing Video Generation for World Complexity"
Official ByteDance pages:
Access and Availability
Seedance models are available through several distribution channels:
ByteDance platforms (China): Doubao (豆包), Jimeng (即梦 / Dreamina), Volcano Engine Ark. These remain the primary distribution channels for users in China.
International access: BytePlus API (ByteDance's international developer platform) and third-party platforms. As of mid-2026, direct API access through ByteDance has seen regional restrictions following IP-related disputes with major studios. Third-party platforms, including our own, remain the most reliable international access path.
On our platform: Seedance 1.0, 1.5 Pro, and 2.0 (Standard and Fast) are available across text-to-video, image-to-video, and reference-to-video generation modes. Check current availability in the generator.
How Seedance Compares to Competing Families
vs. Google Veo (3.0, 3.1)
Veo's strengths: Native 4K output at launch, exceptional audio synchronization precision particularly for dense dialogue, Google ecosystem integration (Vertex AI, Google Cloud), longer maximum clip duration (up to 2 minutes in some configurations).
Seedance's strengths: Stronger physical realism and motion plausibility per blind user comparisons, richer multimodal reference control (Seedance 2.0's 9 images + 3 videos + 3 audio files versus Veo's more limited reference surface), broader language and dialect support for lip sync (8+ languages including regional Chinese dialects), faster inference speed at equivalent quality levels.
The practical difference: Veo is the better fit for teams prioritizing Google Cloud integration, audio precision for dialogue-heavy content, native 4K output, and longer clip duration. Seedance is the better fit for teams prioritizing physical realism, reference-heavy brand work, and multimodal input flexibility.
vs. Kuaishou Kling (3.0, 3.0 Pro)
Kling's strengths: Longer single-pass generation (up to several minutes versus Seedance's 15 seconds), native 4K output, structured per-shot API with precise cut timing control, strong stylized and cinematic aesthetic output.
Seedance's strengths: Richer multimodal reference input surface (Seedance 2.0), stronger overall physical realism per blind user comparisons, broader audio-visual joint generation language support (Seedance 1.5 Pro and 2.0), faster inference at equivalent quality.
The practical difference: Kling is the better fit for long-format narrative work where clip duration is a binding constraint or native 4K is required. Seedance is the better fit for reference-heavy production work, multilingual audio-visual content, and workflows prioritizing multimodal reference control over clip length.
vs. Runway Gen-4, Gen-4.5
Runway's strengths: Integrated editing workflow with frame-level control, broader creative toolset (inpainting, masking, motion brush), strong ecosystem for post-production integration, instruction-based post-generation editing.
Seedance's strengths: Substantially higher raw generation quality per blind user comparisons, stronger physical realism and motion plausibility, native audio-visual joint generation (Seedance 1.5 Pro and 2.0), faster inference at equivalent quality.
The practical difference: Runway is the better fit for teams prioritizing fine-grained post-generation editing control and integration into existing post-production pipelines. Seedance is the better fit for teams prioritizing raw generation quality, physical realism, and native audio-visual synchronization.
Known Limitations Across the Family
Clip length ceiling at 15 seconds. All current Seedance versions cap at 4 to 15 seconds per generation. Longer content requires chaining. Multi-shot mode within a single generation mitigates some of this, but Kling's multi-minute single-pass generation and Veo's 2-minute clips represent a meaningful advantage for long-format work.
International API access variability. ByteDance paused direct global API access in March 2026 following IP disputes with major studios. Access through third-party platforms remains available, but teams building production pipelines that depend on direct ByteDance API access should verify current regional availability before committing.
Multi-person simultaneous lip sync remains challenging. ByteDance's technical papers acknowledge that multi-person simultaneous lip sync matching is an open problem across the family. Single-character dialogue performs best. Multi-character scenes with simultaneous speaking are better approached with audio reference clips or staggered speaking turns.
Reference input requires curation. Providing 9 reference images does not automatically produce better results than providing 2 well-chosen ones (Seedance 2.0). The model responds to the quality and specificity of reference inputs, not their quantity. Over-referencing or providing conflicting references can reduce output coherence.
No extended clip-level temporal control. Kling 3.0's structured per-shot API provides more precise control over cut timing and shot-to-shot transitions than Seedance's natural language shot labeling. For workflows requiring frame-accurate cut control, Kling's API structure is more direct.
The Progression in Context
The Seedance family's evolution reflects a clear strategic progression:
June 2025 (1.0): Establish motion stability as the baseline. Make it harder to get a bad result. Prove that ByteDance can compete with Google and Kuaishou at the top of the leaderboard.
December 2025 (1.5 Pro): Add native audio-visual joint generation. Solve the lip sync problem at the phonome level. Support multilingual and dialect content without separate dubbing pipelines.
February 2026 (2.0): Unify multimodal reference control. Let brand teams and production workflows use multiple images, video clips, and audio files as inputs simultaneously. Process all modalities together for unified creative decisions.
Each version solves a different production constraint. Each version builds on the stability foundation that 1.0 established. The family is not optimized for a single impressive demo but for the messy, varied, real-world workflows that production teams actually need.
That strategic clarity is what makes Seedance a family rather than a series of disconnected model releases.
Which Seedance Model Should You Use?
If you need fast iteration without audio and 1080p is sufficient → Seedance 1.0
The inference speed (10× faster than baseline approaches) and lower credit cost make it the practical choice for high-volume prompt testing, B-roll generation, product animation from stills, and storyboard previsualization where audio will be added in post.
If you need native audio and multilingual lip sync → Seedance 1.5 Pro
Phoneme-level lip sync across 8+ languages, dual-branch audio-visual joint generation, and reference audio-to-video mode make it the direct path from script to synchronized output for dialogue-driven content, multilingual advertising, and audio-synced social media content.
If you need reference-heavy brand work or multimodal input control → Seedance 2.0
The unified multimodal architecture (up to 9 reference images + 3 video clips + 3 audio files per generation) and strongest overall physical realism per blind user comparisons make it the appropriate model for brand content, product visualization with multiple reference assets, and commercial work requiring multimodal reference consistency.
If you're unsure, start with Seedance 2.0. It represents the family's current production ceiling and supports the broadest range of workflows. Use Fast for speed, use Standard for quality.
Frequently Asked Questions
What is the main difference between Seedance 1.0, 1.5 Pro, and 2.0? Seedance 1.0 established motion stability and inference speed as the baseline with silent 1080p generation. Seedance 1.5 Pro added native audio-visual joint generation with phoneme-level lip sync across 8+ languages. Seedance 2.0 introduced unified multimodal reference control (9 images + 3 videos + 3 audio files) with native resolutions of 480p/720p and platform super-resolution to 1080p. Each version builds on the motion stability foundation while solving a different production constraint.
Which version should I use? Use 1.0 for fast iteration without audio. Use 1.5 Pro for multilingual dialogue and audio-visual joint generation. Use 2.0 for reference-heavy brand work and multimodal input control. If unsure, start with 2.0 because it supports the broadest range of workflows.
Does Seedance generate audio automatically? Seedance 1.0 does not generate audio (silent generation). Seedance 1.5 Pro and 2.0 generate native audio (dialogue, sound effects, ambient sound, background music) in the same pass as the video through dual-branch or unified multimodal architectures.
What is the maximum clip length? 2 to 12 seconds for Seedance 1.0 and 1.5 Pro. 4 to 15 seconds for Seedance 2.0. Longer content requires chaining multiple generations. Multi-shot mode within a single generation helps maintain visual consistency across cuts.
Does Seedance support 4K output? Seedance 1.0 and 1.5 Pro generate at 1080p. Seedance 2.0 generates at native 480p/720p with platform-level super-resolution to 1080p per the official technical paper (arXiv:2604.14148). For native 4K output, consider Veo 3.1 or Kling 3.0.
Which languages does Seedance support for lip sync? Seedance 1.5 Pro and 2.0 support phoneme-level lip sync across Mandarin Chinese, English, Japanese, Korean, Spanish, Indonesian, and regional Chinese dialects including Sichuanese and Cantonese. Mandarin and English produce the most consistent results; other languages perform well on short phrases with occasional drift on longer sentences.
Is Seedance available outside China? Yes, through third-party platforms. Direct API access through ByteDance's BytePlus platform has seen regional restrictions since March 2026 following IP-related disputes. Third-party platforms, including our own, remain the most reliable international access path as of mid-2026.
How does Seedance compare to Veo and Kling? Veo leads on audio synchronization precision, native 4K output, and Google Cloud integration. Kling leads on single-pass clip length (up to several minutes) and native 4K. Seedance leads on multimodal reference control (2.0), physical realism per blind user comparisons, and broader multilingual lip sync support (1.5 Pro and 2.0). The best choice depends on the specific workflow requirements.
The Seedance family represents ByteDance's systematic approach to production-scale video generation. Motion stability as the baseline, audio-visual joint generation for narrative content, and unified multimodal reference control for brand work.
