Seedance 2.5: 30-Second Video, 50 References & Native 4K

Developer: ByteDance Seed Lab · Released: July 31, 2026 · Previewed: June 23, 2026 (FORCE 2026) · Official page: Seedance · Model ID: dreamina-seedance-2-5-260628

---

Seedance 2.5 is the current flagship in the Seedance family, introducing 30-second single-pass video generation, native 4K output, and region-level editing. For the full family overview and version comparison, see the Seedance family guide. For a deep dive into the previous generation, see the Seedance 2.0 guide.

What Is Seedance 2.5?

Seedance 2.5 is ByteDance's third-generation AI video generation model, officially released on July 31, 2026, following a preview at the company's FORCE 2026 conference on June 23. It is the first Seedance model to generate native 30-second video in a single pass and the first to deliver native 4K output.

The release represents a structural shift in what AI video generation can cover. Seedance 2.0 demonstrated that AI could produce video that looked convincingly real; Seedance 2.5 extends that realism into scenes long enough to carry a beginning, a middle, and an end within a single generation. The 15-second ceiling of 2.0 required creators to stitch clips together and manage continuity breaks between them. At 30 seconds, a product scene, a short dialogue beat, or a narrative establishing sequence can now live inside one generation.

Two other changes compound the effect: the multimodal reference budget more than triples from 15 combined inputs to 50, and editing capabilities move from clip-level to region-level, so targeted changes no longer require regenerating the entire clip.

---

What Changed from Seedance 2.0

The three headline changes in Seedance 2.5 each solve a different constraint that limited production work in 2.0:

30-second generation removes the stitch problem. The 15-second limit in 2.0 meant that any scene longer than that had to be assembled from multiple generations. That assembly introduced continuity risks: lighting could shift between clips, character identity could drift slightly, camera logic could break at the join. Seedance 2.5 eliminates those risks for scenes up to 30 seconds by generating them in a single pass. ByteDance reports that the model maintains audiovisual consistency across the full duration, where earlier models tended to show visual drift after the 6–8 second mark.

50 reference inputs change what brand work looks like. Seedance 2.0's ceiling of 9 reference images, 3 video clips, and 3 audio files was already the most generous multimodal input surface available at its launch. Seedance 2.5 expands that to up to 30 reference images, 10 video clips, and 10 audio clips—50 total. For brand teams and production studios that need to maintain complex character wardrobes, consistent environments, specific visual tones, and precise audio references across a batch of video variants, the expanded input budget means more of those constraints can be expressed in a single generation call rather than across multiple passes.

Region-level editing changes the revision workflow. In Seedance 2.0, editing was primarily an input problem: refine the prompt, adjust reference inputs, regenerate the clip. Seedance 2.5 introduces the ability to modify a specific region of a generated clip—changing a character detail, swapping an object, or adjusting a scene element—without rebuilding the whole video. Timestamp-level control extends this further, allowing targeted edits at specific moments in the generation.

---

Core Capabilities

30-Second Native Generation

Seedance 2.5 generates video up to 30 seconds in a single forward pass. ByteDance's official positioning describes this as single-pass output: the model maintains one continuous generation context across the full duration rather than chaining shorter segments internally.

For content requiring coherent character identity, consistent lighting, and stable camera logic across a longer duration, this changes the practical ceiling. A 30-second brand film, a product demonstration with a beginning and end, or a narrative scene with a clear arc can now be generated and evaluated as a unit rather than assembled from parts.

For content longer than 30 seconds, Seedance 2.5 supports multi-round extension: generate the initial clip, then extend it in subsequent passes. The model maintains visual and audio continuity across rounds, reducing the visual drift that typically appears at extension boundaries.

50-Reference Multimodal Input

The reference input surface in Seedance 2.5 accepts up to 50 multimodal assets per generation:

• Up to 30 reference images: character identity, costume detail, environment anchoring, prop reference, visual style
• Up to 10 reference video clips: camera language, motion rhythm, action style, pacing reference
• Up to 10 reference audio clips: voice character, music tone, ambient sound, audio style reference

These inputs are processed jointly alongside the text prompt—the same unified architecture introduced in Seedance 2.0, now with a significantly larger reference capacity. The model reasons over all inputs together before generating the first frame, rather than treating them as independent signals mixed afterward.

The expanded budget is most useful when a brief has multiple simultaneous constraints: a specific character, a defined environment, a required visual tone, a specific camera style, and a matching audio reference. In 2.0, trade-offs between competing inputs were necessary. In 2.5, more of those constraints can be expressed directly.

Native 4K Output

Seedance 2.5 generates natively at 4K resolution (3840×2160) with 10-bit color depth. This is a significant change from Seedance 2.0, which generated at native 480p and 720p with platform-level super-resolution processing the output to 1080p.

The practical difference matters for content that will appear at large format, in high-resolution broadcast, or in production pipelines that require source-quality output for downstream processing. The super-resolution path in 2.0 was indistinguishable for web and social media applications; native 4K removes the uncertainty entirely for any use case where resolution headroom matters.

Region-Level and Timestamp Editing

Seedance 2.5 introduces targeted editing capabilities that were not available in the 2.0 workflow:

Region-level editing: Modify a specific area of a generated clip without rebuilding the whole video. Change a character's clothing, swap a prop, adjust a background element—while preserving the camera movement, lighting direction, motion continuity, and scene style of the original generation.

Timestamp-level control: Apply edits to specific moments in the clip. Target the beginning of the scene, a mid-clip beat, or the final seconds without affecting the surrounding content.

These capabilities push the Seedance 2.5 workflow closer to a revision cycle and further from a regeneration cycle. For production work where one element needs to change but the rest of the generation is correct, this is the difference between a targeted fix and starting over.

Additional Generation Features

Green screen mode: Generate video with a clean background suitable for compositing, giving visual effects pipelines a direct extraction surface without a separate keying step.

Camera perspective control: Define camera angle, height, and orientation as parameters rather than relying solely on the text prompt. This provides more predictable control over shot composition, particularly for product visualization and architectural content.

Reference-based editing: Use an existing video as an edit reference so the generation adopts the motion rhythm, pacing, or visual treatment of the reference clip while applying changes specified in the prompt.

---

Generation Modes

Seedance 2.5 supports the same three generation modes as Seedance 2.0—text-to-video, image-to-video, and reference-to-video—with one material change across all three: the maximum clip length doubles from 15 to 30 seconds per generation call. For a full description of how each mode works, see the Seedance 2.0 guide.

The 30-second extension changes the creative scope within each mode in a specific way:

Text-to-video: A prompt can now describe a scene with a beginning, a transition, and a resolution without requiring the creator to split that arc across multiple generations. A 30-second brand narrative, a product reveal with setup and payoff, or a character-driven sequence with a clear internal beat—these now live inside one generation call rather than being assembled from shorter segments.

Image-to-video: The source image still anchors frame one; the model now generates up to 29 subsequent seconds from that starting state. Scenes that previously required two chained clips—a character entering and settling into a space, for example—can be generated as a single continuous output.

Reference-to-video: The same unified multimodal processing introduced in 2.0 now operates across a 50-reference input surface (30 images, 10 video clips, 10 audio clips) instead of 15 total. More constraints can be expressed in a single generation call; fewer trade-offs between competing reference inputs are necessary.

---

Output Specifications

Source: ByteDance official release · BytePlus ModelArk documentation · as of August 2026

---

Performance and Benchmarks

Seedance 2.5 launched on July 31, 2026. As of early August 2026, no independent third-party benchmark results are available—performance data at this stage reflects ByteDance's official launch claims and preview materials, not independent evaluation.

ByteDance's reported improvements over Seedance 2.0 (official claims):

• Approximately 20% improvement in prompt adherence
• Measurable gains in motion stability and frame-to-frame coherence
• Improved shot transition quality in multi-shot generations
• Audiovisual consistency maintained across the full 30-second generation window, where earlier models tended to show visual drift after the 6–8 second mark
• More refined phoneme-level lip sync and synchronized dialogue output

For context: Seedance 2.0 held the top position on the Artificial Analysis Video Arena leaderboard across text-to-video and image-to-video categories (with audio) through much of early 2026. Seedance 2.5 is positioned as a generational step above that baseline.

Independent benchmark results are expected as the model accumulates broader usage and third-party testing access opens. This section will be updated when that data is available.

---

Accessing Seedance 2.5

ByteDance consumer platforms: Dreamina and Doubao Pro are the primary access surfaces, and the first platforms to receive the full Seedance 2.5 feature set.

BytePlus ModelArk API: Seedance 2.5 documentation and pricing are published in the ModelArk console. Broad API access is rolling out in phases; verify your account's regional permissions before building production pipelines against this endpoint.

API pricing (BytePlus ModelArk, as of launch):

• Without video input: $10.70 per million tokens
• With video input: $6.40 per million tokens

Actual generation costs vary based on output resolution and clip duration.

On our platform: Seedance 2.5 is coming soon. Generate with Seedance 2.0 now across text-to-video, image-to-video, and reference-to-video modes.

Note on regional access: Direct ByteDance API access has seen restrictions for some international regions since March 2026 following IP-related disputes with major studios. Third-party platforms remain the most reliable international access path for many regions.

---

Use Cases

Long-scene narrative and campaign production The 30-second ceiling changes what a single generation can contain. A campaign spot with an opening, a product moment, and a closing beat; a character scene that establishes, builds, and resolves; a testimonial that includes context and response—these can now be generated and reviewed as a complete unit, not assembled from parts. The revision workflow changes accordingly: region-level editing lets you fix a specific element in the generated clip without rebuilding the whole scene.

Complex brand briefs with many simultaneous constraints The 50-reference input budget is most valuable when a brief has too many constraints for 15 inputs to express cleanly. A brief that requires a specific character, a defined environment, a costume from a brand lookbook, a camera style extracted from a reference reel, and an audio character matching a previous campaign can now be expressed directly in a single generation call. Fewer constraints get rounded off; fewer passes are needed to get a usable result.

Long-format content via multi-round extension For content that needs to run beyond 30 seconds—a two-minute explainer, a longer brand film, an extended narrative arc—multi-round extension provides a structured path. Generate the first 30-second segment, then extend it in subsequent passes. The model maintains visual continuity and audio consistency across extension rounds, which is the practical problem with naive clip-chaining: the join point is usually where the illusion breaks. Multi-round extension is designed to manage that boundary.

Targeted revision without regeneration Region-level editing represents a different relationship with the generation output. When a first pass is directionally correct but has one wrong element—the wrong expression on the character, a distracting prop in the background, a wardrobe detail that doesn't match the brief—you can target that region specifically. The camera movement, the lighting, and the rest of the frame stay exactly as generated. This is most useful in reference-heavy production work where the overall composition is valuable but one detail needs correction.

Compositing and effects pipelines Green screen mode generates video with a clean background ready for compositing directly. For visual effects workflows that need an AI-generated performance or environment element without a separate keying step, this removes one stage from the pipeline.

---

How Seedance 2.5 Compares

vs. Seedance 2.0

The clearest framing is workflow coverage, not spec deltas in isolation. Seedance 2.0 is designed for short-clip, reference-rich brand work with a ceiling of 15 seconds and 15 total reference inputs. Seedance 2.5 extends that into long-scene production: 30-second single-pass generation, 50 reference inputs, native 4K, and a revision-cycle editing workflow instead of clip-level regeneration.

For a full side-by-side breakdown of what changed at each specification level, see the Seedance family version comparison.

For short-clip brand work with an established 2.0 reference workflow, Seedance 2.0 remains a valid production option with lower per-generation cost. Seedance 2.5 is the better fit when clip length, output resolution, or post-generation editing are binding constraints.

vs. Google Veo 3.1

Veo 3.1 offers competitive audio synchronization precision, particularly for dialogue-heavy content, and strong Google Cloud integration via Vertex AI. Seedance 2.5 now matches Veo 3.1 on native 4K output and exceeds the 15-second ceiling of prior Seedance models with 30-second single-pass generation.

The practical difference centers on reference control. Seedance 2.5's 50-reference input surface is substantially more flexible than Veo 3.1's reference support, making it the better fit for production work with complex brand constraints. Veo 3.1 remains the better choice for teams prioritizing Google Cloud pipeline integration and dialogue-focused audio quality.

vs. Kling 3.0

Kling 3.0 supports longer single-pass generation than Seedance 2.5 (up to several minutes), native 4K output, and a structured per-shot API with precise cut-timing control. Seedance 2.5 closes the 4K gap entirely and provides significantly richer multimodal reference support—50 references versus Kling's more limited reference surface.

For long-format narrative content where single-pass duration beyond 30 seconds is a hard requirement, Kling 3.0 retains a meaningful advantage. For reference-heavy brand work and production workflows depending on complex multimodal input control, Seedance 2.5 offers the richer input surface.

---

Known Limitations

No independent third-party benchmark data yet. Seedance 2.5 launched July 31, 2026. All performance claims are based on ByteDance's official release materials. Independent evaluation is expected as the model accumulates broader usage.

API access is rolling out in phases and regionally restricted in some markets. Full broad access via BytePlus ModelArk is not yet universally available. Direct ByteDance API access has been restricted for some international regions since March 2026. Verify your account's regional permissions before building production pipelines; third-party platform access remains available for most international users.

Reference quality matters more than reference quantity. The 50-asset ceiling raises the ceiling; it does not automatically raise output quality. A well-chosen set of five reference images that express a coherent creative direction will typically outperform fifty loosely related assets. Conflicting references—different character identities, competing visual styles—reduce output coherence. Curate for specificity and consistency, not volume.

Multi-person simultaneous dialogue is still an open problem. Single-character phoneme-level lip sync in Mandarin and English is the most reliable configuration across the Seedance family. Scenes with multiple characters speaking at the same time remain technically difficult. Audio reference clips provide more predictable results than text-only dialogue direction in those scenarios.

---

Frequently Asked Questions

When was Seedance 2.5 released? Officially on July 31, 2026, following a preview at ByteDance's FORCE 2026 conference on June 23, 2026.

Is Seedance 2.5 available on VioEvo? Coming soon. Generate with Seedance 2.0 now using the generator above.

What is the biggest change from Seedance 2.0? Three changes define the upgrade: single-pass generation extended to 30 seconds; reference input capacity expanded to 50 multimodal assets; and region-level editing capability for targeted modifications without regenerating the full clip.

Does Seedance 2.5 support native 4K output? Yes. 3840×2160 (4K), 10-bit color depth, generated natively—not via super-resolution upscaling.

Is the 30-second generation actually single-pass? ByteDance's official positioning describes it as single-pass output. For content beyond 30 seconds, multi-round extension is supported—the model maintains continuity across extension rounds.

What does "50 multimodal references" mean? Up to 30 reference images, 10 reference video clips, and 10 reference audio clips—50 total—alongside a text prompt in a single generation request. These are processed jointly, not sequentially.

Can I use Seedance 2.5 to edit existing video? Region-level editing applies to video generated by the model: define a region and describe the change, and the model applies the edit while preserving surrounding motion, lighting, and continuity. Reference-based editing allows stylistic modifications guided by a reference clip.

Where can I access Seedance 2.5? Dreamina, Doubao Pro, and BytePlus ModelArk API. API access is rolling out in phases. Our platform integration is coming soon.

What is the API pricing? BytePlus ModelArk: $10.70 per million tokens without video input; $6.40 per million tokens with video input. Costs vary by output resolution and clip duration. Verify current pricing in the BytePlus ModelArk console.

---

Seedance 2.5 is coming to our platform. Generate with Seedance 2.0 now while you wait.