Seedance 2.5 Extend

Seedance 2.5 Extend — AI Video Extension (Video-to-Video)

What is Seedance 2.5 Extend?

Seedance 2.5 Extend, from ByteDance Seed, continues a video you already have. Give it a source clip, choose a direction — forward to carry on after the last frame, or backward to show what happened before the first — and describe what happens in the new section, and it renders fresh footage that keeps the clip's subject, motion, look and aspect ratio. It is the continuation half of the Seedance 2.5 audio-video family, built for long-form, reference-controlled storytelling. The output is the new footage only: a forward extension begins on the source's last frame, so you join the two clips together yourself.

Key Features

• •Forward or backward video extension from a plain-text description of the new section (source clip 2–30 s).
• •Set the new footage length (4–30 s) or let the model choose automatically.
• •Keeps the source's aspect ratio and look, with native audio continued in the clip's own style.
• •Up to 30 reference images, 9 extra reference videos and reference audio for character, motion and style consistency (cited @Image 1 / @Video 2 / @Audio 1).
• •480p, 720p and 1080p output at 24 fps (10-bit colour at 1080p), with optional last-frame return for chaining shots.

Best Use Cases

Extend an approved shot that ran a beat short, build a longer sequence through multiple forward rounds, add an establishing lead-in with a backward extension, or chain clips by feeding the returned last frame into the next generation. Because the new footage inherits the source's shape and style, it suits lengthening ads, trailers and social cuts without re-rendering the original. In testing, a 6-second forward extension of a tin-robot workshop clip carried the same bench, colourful blocks, window light and marching motion into the new section and read instantly.

Prompt Tips and Output Quality

Name the clip as @Video 1, pick a direction, and describe one clear beat — the motion, any new element and the ambience. Keep real human faces out of the source; they are declined. Animated, stylised, product and landscape footage extends most reliably, and reference images keep a continued subject on-model.

How to Use Seedance 2.5 Extend

Seedance 2.5 Extend lengthens an existing clip instead of generating a new one. You supply a source video, a direction and a short description of the new section; it returns only the added footage, which you join to the original.

Step by step

1. •Upload the clip to extend as Video (@Video 1): 2–30 s, mp4 or mov, 480p–4K, up to 200 MB. Keep real human faces out of the source — they are declined.
2. •Set Direction: forward continues the clip after its last frame; backward adds what happened before its first frame. The chosen direction is written into your prompt automatically.
3. •Write the Prompt describing what happens next — one clear beat, the motion and any new element. Cite extra inputs as @Image 1, @Video 2 or @Audio 1.
4. •Choose Duration (length of the new footage, 4–30 s, or -1 for automatic) and Resolution.

Recommended Settings

• •Resolution: 480p for drafts and fast iteration, 720p for final renders, 1080p (10-bit) for highest fidelity.
• •Duration: set 5–6 s for a single clean beat; use -1 to let the model match the source's pace.
• •Generate Audio: leave on to carry the soundtrack across the join; turn off for a silent clip.
• •Return Last Frame: enable to chain this extension into the next shot via its last frame.
• •Reference Images / Videos: add to keep a continued character, motion or style on-model.

In testing, a forward 6-second extension at 720p of a tin-robot workshop clip kept the source's 16:9 ratio, 24 fps and native audio, and continued the same bench, colourful blocks, window light and marching motion without a visible seam — the tin robot reached the blocks and toppled the stack.

Remember the output is the new footage only — stitch it to the original to see the full, longer clip. An extension is billed for the source clip on top of the new footage, so it costs more than its own length alone.

Examples to try

Each example below ran end to end at the settings shown — a 6-second forward extension at 720p of the same tin-robot workshop clip — and carried the source's 16:9 shape, 24 fps, stiff marching motion and native workshop audio into the new footage. Remember the output is the new section only, so stitch it onto your clip to see the longer shot.

• •Forward continuation (the default). Prompt: "The wind-up tin robot keeps marching to the end of the bench, reaches the row of colourful wooden building blocks and bumps the stack with its chest; the blocks topple and tumble across the workshop bench as the robot's legs keep ticking. Same stiff clockwork gait, same bench and window light, natural workshop ambience with the clatter of falling blocks." The robot reaches the blocks and scatters them while its gait and the bench light carry on without a seam. Describe one clear beat and set Duration to 5–6 s for a single clean action.

Reference-guided continuation

This card adds one reference image. The model takes up to 30 reference images (cited @Image 1 onward), plus optional reference videos (@Video 2) and audios (@Audio 1); upload a clean reference on a plain background and name it in the prompt to keep a new or continued subject on-model.

• •Reference-guided continuation. Prompt: "A small yellow tin wind-up toy duck, matching @Image 1, waddles into frame from the left as the tin robot keeps marching; the two toys meet beside the row of colourful wooden blocks and pause nose to nose on the sunlit workshop bench. Keep the robot's stiff clockwork gait, the same bench, window light and natural workshop ambience." The duck arrives matching the reference's yellow tin body and orange beak, so the new character stays on-model. Keep the reference uncluttered for the cleanest match.

Further prompts this model is built for: "Continue forward: the robot pivots at the blocks and marches back the way it came," or "A wind-up mouse from @Image 1 scurries past the blocks as the robot keeps ticking."

Does the output include the original clip?

No — it is the new footage only; a forward extension starts on the source's last frame, so join them yourself.

Can I extend a video of a person?

No — real human faces in the source are declined; use animated, stylised, product or landscape footage.

Can I set the length?

Yes — 4–30 seconds, or -1 to let the model choose.

Does it keep the audio?

Yes — native audio continues in the clip's own style unless you turn Generate Audio off.

Can I change the aspect ratio?

No — the output always matches the source clip's shape.