Generate 4 to 30 second clips from text, an image, or multimodal references. Seedance 2.5 raises the reference ceiling to 50 assets, adds MOV output, and responds to integer-second timestamps. What it does not add is 4K.
Seedance 2.5 is selected by default. Available modes, controls, and the point estimate appear in the workspace and can change as the product evolves.
Generated content will appear here
Three reserved demo slots connect generation mode, aspect ratio, and prompt intent. Real videos and posters will replace these clearly marked placeholders after review.
This workspace: up to 1080p
A landscape text-to-video example for continuous camera movement, product consistency, and synchronized ambience across a full 30-second arc.
One continuous product story with a one-line overview, integer-second time blocks, and a closing block of details that must hold to the end.
This workspace: up to 1080p
A mobile-first image-to-video example that uses two approved stills to pin identity and location, then directs only the movement.
Keep the subject from Image 1 inside the room from Image 2 while the camera makes one slow lateral move.
This workspace: up to 1080p
A text-to-video example that uses the official negative-control syntax so dialogue, effects, and music are not forced into the same layer.
Aligned performance, camera, and ambience with rain and train effects retained, and subtitles and background music explicitly excluded.
Facts reviewed:
The official tables describe the model. This page runs one platform route. The two boundaries are far apart on references and identical on resolution, so they are listed separately rather than blended into one list.
The official ceiling sits well above what this page exposes. Fifty reference assets are a model capability rather than a control here: the reference route of this workspace currently accepts no image, video, or audio upload, and image-to-video accepts two images. Resolution is the one row where the two boundaries meet, because the official ceiling is 1080p at 10-bit and this workspace also stops at 1080p. Editing, extension, and last-frame return are documented for the model and have no control on this page.
Seedance 2.5 is a ByteDance Seed video model registered by Volcano Engine as doubao-seedance-2-5-260628 and released on July 31, 2026. It is a later generation than Seedance 2.0 rather than a different family, and ByteDance’s own positioning is deliberately measured: the model page describes it as built for 30-second storytelling with precise reference control and powerful editing, and the official English documentation states plainly that compared with Seedance 2.0 it is not a cross-generational leap in the way 2.0 was compared with 1.5. That sentence matters when competitors describe this model as a generational step change.
The published output specification is where most coverage goes wrong. Seedance 2.5 renders 480p and 720p at 8-bit and 1080p at 10-bit, at 24 fps, and it stops there. The 4K at 10-bit entry appears only in the Seedance 2.0 column of the official capability matrix, so a claim that this model outputs 4K is a misattribution of the previous generation’s specification. Clips run from 4 to 30 seconds and can be extended twice. Output is MP4 or MOV, and the MOV path is new here: H.264 video with yuv444p chroma sampling and PCM audio, which the official difference list recommends specifically for extension and editing work because it holds colour, brightness, and audio-visual consistency better than MP4 across those operations.
The reference ceiling is the largest single change. Seedance 2.5 accepts up to 50 assets in one request, split as 30 images, 10 videos, and 10 audio clips, where the Seedance 2.0 series accepts 15 in total. Reference audio and video may run to 30 seconds combined, against 15 seconds for 2.0. Two capability changes sit inside that number. Audio can now be used as a reference on its own, without being paired to an image or a video, which the 2.0 series does not allow. And multi-view images are now accepted as subject references, where the official guidance previously advised against them.
Prompting follows a published method rather than guesswork. The official structure has four parts: index every reference asset explicitly by upload order and role, write a one-line overview naming subject, location, event, genre, and any special camera move, break the body into timestamped or shot-numbered segments, and close with the details that must hold for the whole clip. Timestamps are written in whole seconds, which is itself a difference: Seedance 2.0 responds to shot numbers and not to timestamps. Negative control is supported for subtitles and for audio, and the audio part can be narrowed further to effects, background music, or dialogue separately. The model carries native prompt support for 11 languages: Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese, and Korean.
One capability deserves its own note because it changes colour handling rather than content: 1080p output at 10-bit. Seedance 2.0 reaches 1080p at 8-bit and reserves 10-bit for its 4K tier, so a 1080p delivery at 10-bit is something this generation adds rather than inherits. The official container guidance pairs that with MOV for extension and editing, which is the combination the documentation points to when a clip will be reworked rather than published as generated.
This page runs the platform route seedance-2.5 rather than a direct account on Volcano Engine, and the two service boundaries differ in ways worth stating plainly. The workspace selector covers 4 to 30 seconds in one-second steps, defaulting to 5 seconds, and offers 480p, 720p, and 1080p, so resolution is the one specification a visitor can match end to end against the official sheet. Reference upload is the sharp divergence: image-to-video accepts two images, and the reference route currently exposes no image, video, or audio upload at all, which is why the official 50-asset ceiling is described above as a model capability rather than a control here. Generated audio is on by default and has no visible switch, and editing, extension, and last-frame return are documented for the model without a corresponding control on this page.
Choose this generation when the work needs a longer single pass, a much larger reference budget, or a delivery container built for later editing rather than for a short draft.
The official matrix registers 30 images, 10 videos, and 10 audio clips for a single request, against 15 assets for the Seedance 2.0 series. Combined reference audio and video may also run to 30 seconds instead of 15.
The official tutorial states that Seedance 2.5 supports video generation from an audio reference alone, with no accompanying image or video. The 2.0 series requires audio to be paired, so this is a new input shape rather than a larger limit.
A single pass covers 4 to 30 seconds and the result can be extended twice, which is the change that moves the model from short shots toward continuous narrative work.
1080p output is 10-bit on this generation, where the 2.0 series reaches 1080p at 8-bit and reserves 10-bit for 4K. The published resolution list ends at 1080p, so pages that advertise 4K for this model are describing the previous generation.
MOV is written as H.264 video with yuv444p chroma sampling and PCM audio, and the official difference list recommends it for extension and editing because it preserves colour, brightness, and audio-visual consistency better than MP4 across those tasks.
Prompts can be timed in whole seconds rather than described only by shot order. The official guidance still warns against controlling frequency with timestamps, such as asking for three head shakes in one second.
Beyond the registered fixed ratios, the model can produce any output ratio between [0.4, 2.5] by controlling the input assets, where the previous generation supports fixed ratios only.
Multi-view images are accepted as subject references on this generation. The official difference list records that the 2.0 series does not recommend them, so this is a reference-shape change rather than a quality claim.
Documented model capabilities for revising elements inside existing footage, continuing a clip forward or backward while holding style and subject, and handing back the closing frame as a usable still.
Choose this generation when the work needs a longer single pass, a much larger reference budget, or a delivery container built for later editing rather than for a short draft.
Test a longer product narrative, camera blocking, sound cues, and transitions before committing to a shoot, using a single pass instead of stitching short clips.
This is the one specification a visitor can match end to end here, because the official ceiling and this workspace both stop at 1080p. The official 10-bit path applies at that resolution.
The model accepts up to 50 references across three modalities. Note that this workspace exposes two images in image-to-video and no upload in the reference route, so a full 50-asset pass has to run on the official endpoint.
MOV output exists for clips that will be extended or edited later, which the official guidance frames as the reason to prefer it over MP4 for that workflow. The editing controls themselves are not exposed on this page.
Explore scenes with several story beats, characters, locations, dialogue, and audio-visual continuity inside one generated 30-second sequence.
Animate approved stills for horizontal, square, or vertical concepts while holding identity, wardrobe, and lighting direction across the full clip.
Each prompt follows the official four-part structure: explicit reference roles, a one-line overview, integer-second segments, and a closing block of details that must hold. Negative control is written at the end rather than mixed into the action.
Uses the full official structure and the full 30-second range, so the model is not asked to compress a complete arc into a few seconds.
One-line overview: a cinematic one-take product story following a single ceramic tea cup from a workshop bench to a quiet window table, natural light, slow continuous camera. 0-6s: macro view of the cup as morning steam catches warm window light, camera low and still. 7-14s: camera pulls back and follows a craftsperson carrying the same cup across the workshop, keeping hand and cup geometry unchanged. 15-22s: camera arcs around a wooden table as the cup is set down beside a folded linen cloth. 23-30s: camera settles into a slow push toward the cup as light shifts across the glaze. Details holding for the entire clip: one cup only, the same hands, one light direction from the left, continuous room tone with a faint ceramic tap. No subtitles. No background music. No logos.
Maps two approved stills to explicit roles, which matches the two-image upload this workspace exposes in image-to-video.
Image 1 sets the subject identity and wardrobe. Image 2 sets the room and the light direction. One-line overview: a restrained portrait sequence in a daylit studio, subject seated, one slow lateral camera move. 0-8s: hold the subject from Image 1 inside the room from Image 2; only a breath and a small shift of the eyes toward camera. 9-18s: camera tracks slowly left as the subject turns slightly and the window light crosses the shoulder. 19-30s: camera settles and the subject returns to stillness. Details holding for the entire clip: facial structure, hairstyle, garment seams, room geometry, and light direction unchanged. Quiet room tone. No cuts. No added people. No on-screen text.
Demonstrates the official negative-control syntax, which can suppress subtitles and background music separately from effects.
One-line overview: a naturalistic two-hander in a rain-lit station cafe, 30 seconds, one location, wardrobe fixed. 0-8s: medium two-shot; the first character says quietly, {you came back}. 9-17s: slow push toward the second character, who answers, {only for tonight}. 18-24s: hold the silence as a train passes outside and reflections cross both faces. 25-30s: camera eases back to the opening framing. Details holding for the entire clip: identical wardrobe, table position, and window geometry in every segment, accurate lip sync, rain and distant train effects retained. No subtitles. No background music. No extra characters.
Keeps one subject and one prop constant across three environments and carries continuity with a repeated sound rather than a match cut.
One-line overview: a 30-second narrative following one red bicycle from a coastal lane to an afternoon market to a bridge at blue hour. 0-9s: sunrise on a quiet coastal lane, rider enters frame, camera tracks alongside at wheel height. 10-19s: match the wheel rotation into the market, keeping rider, bicycle geometry, and travel direction identical. 20-30s: emerge onto the bridge and settle into a wide rear view. Details holding for the entire clip: one bicycle and one rider, matching gear and colours, continuous chain and wheel sound carried across each transition. No logos. No captions. No abrupt cuts.
Plans a mobile-first clip with full-body framing, interface safe space, and a held final frame inside the 30-second range.
One-line overview: a vertical 30-second fashion editorial in a rain-lit arcade, one continuous take, silver outer layer and dark boots. 0-9s: full-body portrait beneath the arcade, rain refracting light on the floor, camera at chest height. 10-19s: the subject walks toward camera, reflections rippling with each step. 20-30s: camera drifts beside the subject and reveals a lit street beyond, holding on a final wide frame. Details holding for the entire clip: face, garment construction, accessories, and walking direction unchanged, subject kept centred with headroom for interface overlays. Restrained fabric motion, rain ambience, distant traffic. No captions. No brand marks.
Uses the multi-view subject reference this generation adds over 2.0, written for the two-image upload this workspace allows.
Image 1 is a front view of the subject. Image 2 is a three-quarter view of the same subject. One-line overview: a slow turntable-style shot that keeps one identity stable while the camera arcs around it. 0-10s: the front view holds as the camera begins a slow clockwise arc. 11-20s: the three-quarter view from Image 2 becomes visible; keep the same facial proportions, hairline, and garment seams. 21-30s: camera completes the arc and stops on a three-quarter rear view. Details holding for the entire clip: one person, one garment, one light source from the upper left, neutral studio background. No camera shake. No text. No additional props.
The four rows the official documentation writes out as differences are reproduced here verbatim in substance, followed by quantitative rows taken from the same official tables that list both models, and by token rates from the published Ark price page.
| Comparison | Seedance 2.5 | Seedance 2.0 |
|---|---|---|
| Official model ID | doubao-seedance-2-5-260628 | doubao-seedance-2-0-260128 |
| Official release date | July 31, 2026 | February 12, 2026 |
| Clip duration | 4–30 seconds, plus two extensions | 4–15 seconds |
| Output resolution | 480p and 720p at 8-bit; 1080p at 10-bit | 480p, 720p, and 1080p at 8-bit; 4K at 10-bit |
| Frame rate | 24 fps | 24 fps |
| Container | MP4, MOV | MP4 |
| Reference assets in one request | 50: 30 images, 10 videos, 10 audio clips | 15: 9 images, 3 videos, 3 audio clips |
| Reference audio and video, combined length | Up to 30 seconds | Up to 15 seconds |
| Audio used as a reference on its own | Supported without an image or a video | Not supported; audio must pair with an image or a video |
| Timestamp control | Responds to integer-second timestamps | Responds to shot numbers, not to timestamps |
| Multi-view images as subject references | Supported | Not recommended |
| Aspect ratio control | Fixed ratios plus any ratio between [0.4, 2.5] by controlling input assets | Fixed output ratios only |
| Reference loading semantics | Tasks are split by whether the input references lock the output properties | No such distinction |
| Token rate, 480p and 720p, no video input | ¥70 per million tokens | ¥46 per million tokens |
| Token rate, 1080p | ¥77 per million tokens (list price) | ¥51 per million tokens |
| Token rate, 4K | Not supported | ¥26 per million tokens |
Volcano Engine prices this model by tokens. Without video input, 480p and 720p output is ¥70 per million tokens and 1080p is ¥77. With video input the rates fall to ¥42 and ¥46. Official per-clip examples for a 5-second 16:9 output list ¥3.36 at 480p, ¥7.56 at 720p, and ¥18.71 at 1080p, and with video input a 5-second 1080p output runs between ¥20.12 and ¥78.25 depending on the length of the source. One dated promotion applies: from 14:00 Beijing time on August 14, 2026 to 14:00 on September 17, 2026, 1080p output is charged at 72 percent of list price, which the documentation puts at roughly ¥2.70 per second at the floor; 480p and 720p are not part of that promotion. Offline inference is not available. All figures here were read from the official price page on September 12, 2026, and this site charges platform points instead, so the live estimate in the workspace is the price that applies here.
Open the workspace to see the current estimateThe workspace quotes points before you submit. Points are set by this site and are not the official token rate.
480p, 720p, and 1080p from 4 to 30 seconds on the current provider route, matching the official range of 4 to 30 seconds. 720p is the default and there is no 4K option.
Number each image, video, and audio clip by upload order and state what it controls: identity, wardrobe, scene, camera rhythm, voice, or ambience. The official guidance warns that showing the mapping only inside the image leads to confused or duplicated subjects.
Name the subject, location, event, genre or style, and any special camera move in one sentence. This is the part the model reads as the task rather than as decoration.
Break the body into whole-second segments and give each one an action, a camera behaviour, and a sound cue. Do not use timestamps to control frequency, and keep one camera movement per segment rather than stacking several.
End the prompt with the constants: camera position, scene continuity, sound, and atmosphere. This is where wardrobe, prop counts, and light direction belong.
Before generating, check that the workspace shows Seedance 2.5 and review the current duration, resolution, aspect ratio, and point estimate. Resolution is the specification that matches the official sheet; reference upload is the one that does not.
ByteDance describes this generation as an improvement rather than a step change. The official English documentation states that compared with Seedance 2.0, Seedance 2.5 is not a cross-generational leap in the way 2.0 was compared with 1.5, and the launch announcement records that room for improvement remains, particularly in the physical plausibility of complex motion and the stability of scenes with several interacting subjects. Treat every generation as review material rather than as a finished shot.
Three documented behaviours change how a result can be used, and all three are narrower than a general quality statement. Video editing output can be up to 0.4 seconds shorter than the input clip because of input frame handling. Supplying a first and last frame with different aspect ratios stretches the picture, so compatible source images are required. Clay-model control is stated to work better on simple geometry, and the official guidance recommends restricting it to simple shapes representing people, objects, or animals.
Task types carry their own constraints and will error rather than degrade when they are violated. All-modal reference tasks require the ratio parameter to be adaptive. Editing tasks require duration to be set to -1. A reference video must be between 4 and 30 seconds. The published error codes for these cases are explicit, which means a rejected request is a configuration problem rather than an output-quality problem.
Access is limited by rate and by balance. The official tables register a maximum of 600 requests per minute for enterprise accounts and 180 for individual accounts, with 10 and 3 concurrent online inference tasks respectively. Activation requires an account balance above ¥200, or a savings plan at the ¥200 tier or above, or a Seedance 2.5 resource pack with remaining quota. Offline inference is not supported.
Price and promotion statements on this page are dated. The token rates, per-clip examples, and the 72 percent promotion on 1080p output were read from the official price page on September 12, 2026, and that promotion ends at 14:00 Beijing time on September 17, 2026. Anything published here needs re-checking against the official page after that date, and the 480p and 720p rates carry no promotion at all.
Upload only material you are authorised to use. Do not impersonate people, mislead viewers, violate privacy or intellectual-property rights, or omit disclosures required for synthetic media. Model access, API quotas, safety rules, and pricing can change; the official ByteDance and Volcano Engine documentation is authoritative for the model, and the live workspace is authoritative for this site.
Official model identity, specifications, capabilities, limits, and pricing are separated from this platform’s controls, points, and availability. Platform statements were reviewed on the date shown in the facts section above.
Open the live workspace, refine your prompt, review the current parameters and point quote, then generate with the selected Seedance 2.5 mode.