Seedance 2.5 has gone live in full, and its official Feishu guide presents the model as a step up from Seedance 2.0 in generation length, controllability, multi-modal references and editing tools. The manual describes that change in practical terms: the workflow is no longer centered on producing a single fragment, but on building a more complete video sequence in one pass and then refining it.
Native generation reaches 30 seconds, with extensions up to 180 seconds
According to the handbook, Seedance 2.5 raises native generation length for a single run to 30 seconds and supports repeated extensions. After generating a clip within 30 seconds, users can continue extending it through multiple rounds, with total length reaching 60 seconds. The update also adds an ultra-long video mode with support for up to 180 seconds.
The official guidance ties those options to different production needs. For ads, short films and product showcases that need a complete narrative arc, the handbook recommends generating a 20- to 30-second one-take sequence directly. For longer storytelling, it suggests locking in a stable 15- to 30-second core segment first so character appearance, style and pacing hold steady before extending the rest. For longer-form material such as tutorials or full short-drama sections, users can switch to the ultra-long mode, though the guide advises planning overall pacing with timestamps in advance to reduce drift in later sections.
The stated benefit is continuity. By relying less on stitching separate clips together, the model is meant to reduce broken camera flow and consistency problems and make it easier to carry a story from setup to development, turn and ending inside one sequence.
Timestamp prompting is positioned as a core control layer
The handbook identifies timestamp-based text control as one of the central features in version 2.5. Prompts can specify timestamps down to the second, allowing users to direct how actions evolve on screen, how camera movement changes and when scene transitions happen.
The practical guidance is fairly specific. Each time window should carry one core action and one camera instruction, rather than several competing requests. The manual says each window should generally run at least 3 seconds, because shorter spans can make execution unstable. Camera movement, sound effects and dialogue can all be written directly into the relevant time slot.
The guide also says users can make targeted changes to selected time ranges after generation, including edits to characters, motion, sound and plot, while preserving continuity before and after the revised section. In the language of the document, timestamp control moves the model away from loosely interpreted output and closer to script-level execution.
Multi-modal references are expanded into a broader control toolkit
Seedance 2.5 also broadens how reference materials are used. The handbook recommends uploading references by priority in the following order: identity images for the main character or product, white model or storyboard images, camera-motion reference video, voice tone, then scene and style references.
It also asks users to state clearly what each reference is for inside the prompt. One image might be used to lock character identity and wardrobe, while a video clip might be used only for camera movement. That structure is designed to make character consistency, spatial control and shot-language replication more dependable across a sequence.
Cleaner defaults, stronger base quality and direct multilingual prompting
Beyond generation length and control methods, the manual lists several other changes. One is cleaner default output, with fewer irrelevant subtitles and less background music, giving editors more room to add text and music in post-production. When needed, prompts can also explicitly request actions such as separating or removing background music while keeping the voice track.
The handbook also says base generation quality has improved across the board. It describes the new output as less visibly AI-generated, with stronger physical texture, better consistency between cuts and more natural handling of complex long-tail actions. On that basis, users are encouraged to attempt more demanding camera motion and multi-character interaction.
Another item is multilingual support. The guide says users can write prompts directly in the target language, without translating first, for mainstream foreign-language scenarios.
The recommended workflow runs from planning to local edits
The handbook lays out a full production flow rather than treating prompting as a single step.
- Planning: break pacing into timestamps first, or prepare white models and multi-panel storyboards in advance to define rhythm and spatial relationships.
- Reference preparation: upload assets by priority and bind each one to a specific purpose in the prompt.
- Prompt writing: combine timestamps, explicit camera direction and reference binding, writing the prompt more like a storyboard script than a static image description.
- Generation: create the 30-second core segment first, then decide whether to extend after checking pacing and consistency.
- Adjustment and refinement: use extension, local editing, viewpoint changes and green-screen replacement tools for precise revisions instead of regenerating the full piece.
- Output: move cleaner source material into post-production packaging.
The logic in that workflow is straightforward: stabilize character identity, visual style and pacing early, then expand outward, rather than aiming for the maximum possible duration from the start.
Prompt rules and scenario mapping
The guide gives a concise set of prompt-writing rules. Prompts should be treated as storyboard scripts rather than static visual descriptions. Timestamps should break the sequence into windows, and each window should focus on one core action. Camera language should be explicit, with examples in the manual including slow push-in, circular follow shot, low angle and stabilized handheld movement. Every reference asset should be tied to a defined role. If cleaner output is the goal, the prompt should also actively restrict random subtitles and extra background music.
The handbook maps those methods to specific use cases. A 30-second one-take ad or short film is paired with timestamps plus core references. Multi-character scenes are paired with multi-person references and explicit identity binding. Complex camera work and spatial blocking are paired with white-model control. Background replacement and effects are paired with green-screen editing. Long-form narrative work is paired with a 30-second segment plus extension, or the ultra-long mode. Overseas-facing content is paired with direct prompting in the target language.
The central theme is tighter control
Across the handbook, the underlying message is not just that clips can run longer. The larger change is that timestamps, white-model planning, multi-reference input and precise post-generation editing are now framed as parts of one connected workflow. In the guide’s own structure, timestamps manage pacing, white models manage space, multiple references manage consistency and precision editing handles localized revisions.
This article is based on the official Seedance 2.5 usage handbook and its coverage of model positioning, core upgrades, timestamp control, white model and green-screen workflows, multi-reference methods and operational suggestions.