⏱️ Lectura: 11 min
ByteDance doubled the maximum length of its AI-generated videos in one leap: Seedance 2.5, unveiled on July 31, 2026, generates clips of up to 30 seconds with synchronized audio in a single pass, double its predecessor.
📑 En este artículo
The leap isn’t just about length. The model now accepts up to 30 images, 10 videos, and 10 audio clips as reference in a single generation, and lets you extend a clip multiple times without losing character or scene consistency. ByteDance’s promise is concrete: moving from generating a clip to completing a full piece.
TL;DR
- ByteDance launched Seedance 2.5 on July 31, 2026, its new video-with-audio generation model.
- The model generates clips of up to 30 seconds in a single pass, double the 15 seconds of Seedance 2.0.
- It supports multi-round extension to produce multi-minute videos with consistent characters and style.
- It accepts up to 30 images, 10 video clips, and 10 audio clips as reference in a single generation.
- It adds editing with timestamp-level control, green screen, camera perspective changes, and reference-based editing.
- It’s already available on Jimeng AI and Doubao Pro; API access is coming soon via BytePlus ModelArk.
- It’s the direct evolution of the unified audio-video architecture that debuted with Seedance 2.0.
Introduction
The Chinese company behind TikTok has spent a year and a half pushing the frontier of AI video generation through its Seed team. With Seedance 2.5, the stated goal is no longer to compete for the most spectacular few-second clip, but for the complete narrative: a scene that opens, develops, turns, and closes without visible cuts.
This matters for any developer or studio building products on video generation: the minimum unit of work stops being a standalone shot and becomes a narrative sequence, and editing tools (green screen, timestamp control, multimodal referencing) start to resemble a real production workflow instead of a novelty generator.
What Happened With Seedance 2.5
On July 31, 2026, ByteDance’s Seed team officially announced the launch of Seedance 2.5, built on the unified joint audio-video generation architecture that debuted with Seedance 2.0. The announcement highlights three areas of improvement: long-form generation with multiple extensions, expanded multimodal referencing, and more precise and stable editing.
In concrete terms, the model can produce up to 30 seconds of video with audio in a single pass (versus 15 seconds for the previous generation) and chain multiple rounds of extension to produce multi-minute content while keeping the same audiovisual language. ByteDance illustrates this with its own example: a continuous shot of a singer getting ready backstage, walking down the backstage hallway, meeting her dancers, and stepping out onto the stage, all as a single continuous take instead of cuts between independent clips.
The second area is multimodal referencing: users can upload up to 30 images, 10 video clips, and 10 audio clips as reference material in a single generation. The model also strengthens specific capabilities such as clay render referencing, motion referencing, and creative style referencing, to support ideas spanning multiple subjects, scenes, and shot changes.
The third area is editing: Seedance 2.5 offers timestamp-level control for targeted audio and video editing, and improves advanced features like green screen, camera perspective changes, and reference-based editing, designed for professional film and advertising workflows.
Context and History
ByteDance, the company behind TikTok and Douyin, runs its generative AI ecosystem through Doubao (its conversational assistant) and Jimeng AI (its visual creation platform). The Seed team is the internal lab responsible for the company’s foundation models, including the Seedance video generation family.
Seedance 2.0 introduced the architecture for jointly generating audio and video in a single model pass, instead of the traditional approach of generating silent video and adding sound afterward with a separate model. That design decision is the foundation Seedance 2.5 builds on: since audio and video aren’t split into two systems, extending a clip means extending both tracks coherently, with no manual resynchronization needed.
The race among text-to-video models (and now reference-to-video models) has seen just over two years of accelerated development, with labs competing on length, character consistency, and motion quality. ByteDance’s emphasis on a single take to tell a complete story marks a different approach compared to models that prioritize short, visually high-fidelity clips without narrative continuity between takes.
Technical Details and Performance of Seedance 2.5
The centerpiece of Seedance 2.5 is multi-round extension: instead of generating an isolated clip and asking a separate process to continue it, the model itself receives the previous video as context and generates the next portion while keeping characters, setting, visual style, and narrative pacing consistent. ByteDance documents this with a reference-to-video (R2V) example: starting from a first clip of a child running with a ball inside a subway car, the model extends the scene by 30 more seconds, with the child getting off at a stop and being chased by the protagonist out onto the street, keeping the same character and the same visual palette.
The following diagram summarizes the flow of a generation with extension: the input data (text and references) feeds the model, which produces a clip, and that clip can be fed back into the same model to produce the next round.
flowchart TD
A["Text prompt"] --> D["Seedance 2.5"]
B["Up to 30 reference images"] --> D
C["Up to 10 reference videos and 10 reference audios"] --> D
D --> E["Clip up to 30s with audio"]
E --> F{"Needs more length?"}
F -- "Yes" --> G["R2V extension from previous clip"]
G --> D
F -- "No" --> H["Final exported video"]The table below compares the main areas between the previous and current generation, as described in ByteDance’s official announcement:
| Feature | Seedance 2.0 | Seedance 2.5 |
|---|---|---|
| Length per generation | Up to 15 seconds | Up to 30 seconds, with multi-round extension |
| Multimodal referencing | Basic image referencing | Up to 30 images, 10 videos, and 10 audios |
| Editing | Limited | Timestamp control, green screen, perspective change |
| Narrative continuity | Independent clips | Stories with setup, development, turn, and resolution |
To structure a prompt for generating from scratch (text to video, no references), the recommended format describes the shot as a continuous camera movement, not as a list of separate events:
T2V Prompt (text to video, single take):
"Drone continuous shot: the camera descends over a craft
market in Oaxaca at dawn. A weaver works on a shawl
while chatting with a buyer. The camera moves in close
on her hands weaving and then rises to reveal the entire
market waking up in the morning light."This type of prompt works because the model interprets the scene as a single camera trajectory with events happening within it, rather than as independent cuts that would later need to be stitched together in editing.
💭 Key point: The difference from generating and then editing is that multi-round extension reuses the previous video itself as context for the model, so the character, lighting, and ambient audio aren’t recalculated from scratch each round.
How to Start Trying It
Seedance 2.5 is already available, with no waitlist, on two ByteDance products:
- Jimeng AI: go to the Jimeng website, choose the Video Generation section, and select Seedance 2.5 in the model selector.
- Doubao Pro: within the Doubao Pro app or website, go to Video Generation and likewise select Seedance 2.5.
Programmatic access still has no public documentation: ByteDance announced that API access is coming soon through BytePlus ModelArk, its model platform for external developers. Until that’s published, there are no official endpoints or parameters for integrating Seedance 2.5 into your own pipeline, so any HTTP request example would be speculative.
What you can prepare today is the design of extension prompts, since the syntax described in the announcement (referencing a previous video by name and requesting continuity of subjects, scene, and style) is likely to carry over to the API once it’s available:
R2V Prompt (extension with video reference):
"Extend the video. Continue from the subjects and setting of
@Video 1 and generate 30 additional seconds, keeping the same
character, the same location, the color palette, and the
ambient sound. The weaver finishes the shawl, hands it to
the buyer, and they both walk toward a coffee stand at the
back of the market."To confirm that a generation actually used Seedance 2.5 and not an earlier version, check the model selector before generating: both Jimeng and Doubao Pro keep previous versions available in parallel, so the model name should appear explicitly in the interface before you submit the prompt.
Impact and Analysis
For small studios and independent creators, multi-round extension reduces a real bottleneck: today, producing a multi-minute video with generative AI means generating short clips separately and then stitching them together by hand, fixing cuts, lighting, and character continuity in a traditional editor. If consistency across rounds holds up in practice, that post-production work is reduced.
The expanded multimodal referencing (30 images, 10 videos, 10 audios) also changes the kind of control the user has. Instead of describing everything in text and hoping the model correctly interprets the tone, palette, or camera style, you can anchor the generation to concrete material: a brand’s logo, a character’s already-recorded voice, or the style of a reference video.
A real limitation: for now, Seedance 2.5 is only available on two platforms operated by ByteDance from China (Jimeng AI and Doubao Pro), with no public API yet. For a team in Latin America wanting to integrate it into their own product, the only option today is to wait for BytePlus ModelArk to open up; there’s no SDK, no per-token pricing, and no published rate limits, which makes it impossible to plan production costs around this model for now.
⚠️ Heads up: Without a public API or published pricing policy, any product plan built around Seedance 2.5 today depends on the Jimeng and Doubao Pro platforms, with the regional and language restrictions that come with that.
What’s Next
The next announced milestone is API access on BytePlus ModelArk, with no exact date confirmed beyond soon. That launch is what will determine whether Seedance 2.5 becomes an integrable component in third-party products (video editing, marketing, e-learning tools) or remains, for now, a consumer tool within ByteDance’s ecosystem.
It also remains to be seen how other generative video labs respond to this bet on length and multi-round extension instead of just short-clip quality, and what moderation and watermarking policies ByteDance will apply when it opens access to external developers outside China.
📖 Summary on Telegram: View summary
Try it yourself: go to Jimeng AI or Doubao Pro, select Seedance 2.5 in the model selector, and generate your first 30-second clip today.
Frequently Asked Questions
What is Seedance 2.5?
It’s the latest video-with-audio generation model from ByteDance’s Seed team, unveiled on July 31, 2026, capable of producing clips of up to 30 seconds in a single pass and extending them across multiple rounds.
How is it different from Seedance 2.0?
It doubles the maximum length per generation (from 15 to 30 seconds), adds multi-round extension to produce multi-minute videos, expands multimodal referencing to 30 images, 10 videos, and 10 audios, and adds editing with timestamp control, green screen, and camera perspective changes.
How can I try it from Latin America?
Today it’s only available within Jimeng AI and Doubao Pro, by selecting Seedance 2.5 in the model selector within the video generation section. It doesn’t require a waitlist, but both platforms are operated from China.
Is it an open-source model?
No. ByteDance didn’t release the model’s weights or code; access is exclusively through its own platforms, with API access announced for the near future via BytePlus ModelArk.
How much does it cost to use?
ByteDance didn’t publish pricing or usage limits for Seedance 2.5 at the time of the announcement. API pricing will depend on what’s published when BytePlus ModelArk opens access to developers.
What kind of material does it accept as reference?
Up to 30 images, 10 video clips, and 10 audio clips in a single generation, with strengthened support for style, motion, and clay-render referencing.
References
- Official Seedance 2.5 announcement: ByteDance’s Seed team blog with full details on capabilities and examples.
- Seedance 2.5 project page: video demos of long-form generation and multimodal referencing capabilities.
- ByteDance on Wikipedia: background on the company and its generative AI product ecosystem.
- Text-to-video models on Wikipedia: general background on the model category Seedance belongs to.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Jesus Loves Austin en Unsplash
0 Comments