Two generations, one vendor, three limits that moved. Two went up. The third went sideways, in the direction that surprises anyone who assumes the newer model is the larger one.
What changed between Seedance 2.0 and Seedance 2.5
ByteDance lists both generations in one capability matrix on its Volcengine Ark documentation, and that page is the only place the numbers sit side by side. The upgrade story reads short: longer single generations, more reference material, one extra output container. The resolution column breaks the pattern, so it stays in view below rather than getting smoothed over. Every figure here comes from a page ByteDance publishes, and the Seedance hub tracks the model line as it changes.
The parameter table, and only the vendor’s own rows
The 2.0 column covers the 2.0 series, since 2.0, 2.0 fast and 2.0 mini share these values apart from the note underneath.
| Parameter | Seedance 2.5 | Seedance 2.0 series |
|---|---|---|
| Model ID | doubao-seedance-2-5-260628 | doubao-seedance-2-0-260128 |
| Duration per generation | 4~30 秒, or -1 to let the model choose | 4~15 秒, or -1 to let the model choose |
| Reference material per request | 50, split as 30 images, 10 videos, 10 audio | 15, split as 9 images, 3 videos, 3 audio |
| Output resolution | 480p (8-bit), 720p (8-bit), 1080p (10-bit) | 480p (8-bit), 720p (8-bit), 1080p (8-bit), 4k (10-bit) |
| Audio as a reference | Supported on its own | Needs an image or a video alongside |
| Output format | mp4, mov | mp4 |
| Output aspect ratio | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, adaptive | the same seven values |
| Draft mode, offline inference | Neither is supported | Neither is supported |
The two smaller 2.0 variants, fast and mini, cut the resolution list to 480p and 720p.
Duration: fifteen seconds becomes thirty
The changelog states it in one line: the duration ceiling moves from fifteen seconds to thirty. The matrix puts the same limit in cells, 4~30 秒 for Seedance 2.5 against 4~15 秒 for the 2.0 series, each with a value of -1 that hands the choice of length back to the model.
Two details get lost in summaries. The floor is four seconds, so a two-second insert sits outside the documented range on either generation. And the ceiling is per generation: the English project page adds that a 2.5 clip can be extended twice, which is how several minutes of footage gets built, while the Chinese launch post says multi-round extension without giving a count. The Seedance API guide covers the task lifecycle those extra calls run through.
Reference material: fifteen items become fifty
Fifty items per request — thirty images, ten videos, ten audio clips. The 2.0 series stops at fifteen: nine, three and three. The changelog splits the jump into three lines: the image cap from nine to thirty, the audio and video count from three to ten each, and the combined running time of reference audio and video from fifteen seconds to thirty.
One capability is new rather than merely larger. Seedance 2.5 accepts audio on its own, so a request can carry a music bed or a voice clip with no image or video beside it. On the 2.0 series the matrix marks audio reference as needing an image or a video alongside.
Resolution, where the older generation keeps a tier
This is the row that trips people up. Seedance 2.5 offers 480p, 720p and 1080p. The 2.0 series offers those three plus 4k, which settles the “4k or 1080p” question: 4k belongs to 2.0, and the current generation tops out at 1080p.
No 2K tier exists to fall back on either. If a comparison table hands Seedance 2.5 a 2K or a 4k output, the vendor’s matrix does not contain one. Inside a tier, 2.5 does change bit depth: its 1080p is 10-bit where 2.0’s is 8-bit. Containers differ too: 2.5 writes mp4 or mov, the 2.0 series mp4, and the vendor describes the mov addition as H.264 video, yuv444p chroma sampling and PCM audio.
Worked example: one prompt, two model IDs
Two prompts the vendor published, each sent to both generations.
Case one: a thirty-second narrative prompt
The launch post pairs a text-to-video prompt with its output clip. The prompt opens 一镜到底手持稳定器跟拍，镜头从红色厚重幕布缝隙缓缓推进，进入暖色后台化妆间 and closes 镜头最终拉远至体育馆全景，呈现满场观众. That is a singer’s backstage-to-stage run in one take, sent to Seedance 2.5 like this:
{
"model": "doubao-seedance-2-5-260628",
"content": [{ "type": "text", "text": "<the prompt above>" }],
"ratio": "16:9",
"duration": 30,
"generate_audio": true
}ByteDance describes the result in one line: the model plays the dressing-room exchange, the backstage corridor, the meeting with the dancers and the walk-on as a single continuous story, rather than one shot of the singer arriving on stage. Change a field and it behaves differently: point the prompt at doubao-seedance-2-0-260128, keep duration at 30, and the value falls outside the documented 4~15 秒 range, so the text has to become two requests of fifteen seconds or less.
Case two: a prompt that needs eighteen reference images
The second published prompt is a concert scene, and it addresses its material inline: 场景@图一，钢琴家参考@图二，大提琴参考@图三，小提琴参考@图四，主唱参考@图五，管弦等其他乐队参考@图六到图十，合唱团参考图十一到图十四，观众席参考@图十五到图十八. Eighteen images, one request. On Seedance 2.5 that fits inside the thirty-image allowance, and the vendor’s description of the output is that the model restores several people’s appearance and voice at once and keeps each subject stable in a crowded frame. Take the same prompt to a 2.0 model and the image allowance is nine, so half the references go.
Which generation to reach for
Match the model to the limit you actually hit. Narrative work that runs long needs 2.5, the only one of the two that documents thirty seconds in a single call. Scenes built from a crowd of referenced subjects, or from a music bed, also need 2.5 for the larger allowance and the audio-only path, which the how-to-use guide shows in practice. If you need 4k output, the older model lists it, and recency does not change that. Plain text prompts for short clips are where the two converge, same aspect-ratio list and the same draft-mode and offline-inference limits, so the difference narrows to duration and reference capacity.
What the vendor has not published
Three gaps worth naming rather than filling. The capability matrix carries no frame rate, and no page consulted here states one, so none appears above. Reference-video input constraints — per-clip duration, clip count, combined length, file-size cap — live on the vendor’s separate task-creation page. And no performance delta between the generations is published: the 2.0 project page shows a vendor-run internal benchmark, the 2.5 project page shows none, and neither gives a percentage comparing them.
Frequently asked questions
How long can one generation be? Four to thirty seconds on Seedance 2.5, or -1 to let the model choose inside that range. The 2.0 series stops at fifteen.
How much reference material does each accept? Fifty items on 2.5 — thirty images, ten videos, ten audio clips. Fifteen on the 2.0 series — nine, three and three.
Does Seedance 2.5 output 4k? No. Its published tiers are 480p, 720p and 1080p. The 4k tier belongs to the 2.0 series.
Can I send audio with no image or video? On Seedance 2.5, yes. On the 2.0 series an audio reference needs an image or a video alongside it.
Three rows, two directions: the newer generation is longer and hungrier for reference material, and the older one still holds the top resolution tier.