Seedance

Seedance 2.5

The newer of the two, and the one with the most room: thirty reference images, clips up to thirty seconds, and several shots inside one generation. It is the one Lorvara sizes a video prompt against when a project has not named a model of its own.

References. Up to thirty pictures, addressed by role (characters, props, locations) exactly as the rest of the roster is. Write them by name in the prompt and Lorvara renders the model’s own token at send.

Multi-shot. A single generation can carry up to six shots, inside a total of thirty seconds. It reaches the model as one prompt with the shot structure written into it (the endpoint has no separate shots parameter), and the cut between shots is the model’s, not an edit Lorvara makes. There is no element list on this endpoint, which is a different thing from a reference it may decline.

Duration, size and sound. Any whole number of seconds from 4 to 30. 480p, 720p or 1080p. Aspect ratios from 21:9 through to 9:16. Audio is generated by default and can be turned off.

No start or end frame. The endpoint declares first- and last-frame roles, but Lorvara has not exercised them, so the controls are not offered. Claiming a frame path that has not been measured would put it in the picker and refuse it at Generate.

“Realistic face refs must be registered”

All three Seedance cards carry this note, and it is worth reading carefully, because it is a note and not a refusal.

The Seedance family declines a photoreal face in a reference image at submit time, unless it belongs to a character you have registered on the brief’s Cast. A stylized character passes. That is measured, not assumed. Lorvara cannot tell the two apart in your browser and does not try: the references go through exactly as they do on any other model, and if Seedance declines one you get its own message, and your credits come back.

So the note is there to set an expectation before you spend, not to stop you.

Where it runs

Every Seedance model is served through our model aggregator, and processed in Malaysia or China (Beijing). Each card’s rights line names the company that serves it.

The aggregator routes each request dynamically between its infrastructure in Malaysia and its facility in Beijing, in mainland China. There is no primary facility and no failover, and nothing on the request can pin which one runs it, so both are named, because promising one would be a claim we cannot keep.

This matters if you are sending reference pictures of people. See section 9 of the privacy policy for what that means for the material you upload.

Seedance 2.0

The established version.

References. Up to nine pictures, through the reference-to-video path.

Duration, size and sound. 4 to 15 seconds, 480p / 720p / 1080p, the same range of aspect ratios, audio on by default.

Start and end frames. Unlike 2.5, Seedance 2.0 takes both, so it is the one to reach for when a shot has to begin or end on a specific image.

The same face note. Realistic face refs must be registered means the same thing here as it does on 2.5, for the same reason: it is the family’s behavior, not a Lorvara rule. 2.0 is served through our model aggregator as well, so Where it runs applies to it too.

No multi-shot. Several shots in one generation is 2.5’s, not 2.0’s.

Seedance 2.0 Mini

The lighter member of the family. It is the same model to work with as 2.0 (the same references, the same durations, the same aspect ratios, the same audio) at smaller sizes.

References. Up to nine pictures, as 2.0.

Duration, size and sound. 4 to 15 seconds. 480p or 720p, and no 1080p: that is the one thing that distinguishes it, so a shot you need at full size belongs on 2.0 or 2.5. The same range of aspect ratios, audio on by default.

No start or end frame, and no multi-shot. Start and end frames are 2.0’s alone; several shots in one generation are 2.5’s alone.

The same face note, and the same place. Realistic face refs must be registered means what it means on the other two, and Mini is served through the same aggregator, so Where it runs applies here as well.

Prompt length on all three

None of them states a prompt limit of its own. Our model aggregator recommends about 1,000 English words (roughly 6,000 characters), and that is the length Lorvara’s Prompt Agent writes to; a probe of 60,000 characters met no length rule, so Lorvara refuses above that as a safety stop, with a line break counting as two. See The model picker.