Seedance 2.5 is coming, and it's aimed at everything AI video gets wrong

Right now, over a quarter of video generation traffic on Melius is routed to Seedance 2.0. Initially, Seedance wasn't heavily used, and didn't even cover a tenth of our generations in regards to videos. It just quietly became one of the models people here reach for when they need a quality video to post, and its share has climbed nearly every month since our inception in spring. When ByteDance announced Seedance 2.5, it was not one more model to go and evaluate. It's an upgrade to something a lot of users are already excited about and are still pushing the boundaries of.

This piece is two things: what our own generation data says about how people use Seedance right now, and a read on what 2.5 promises. The usage numbers here are ours, and they are pulled directly from our data. Seedance 2.5 is not public yet, so its specs are based on what info ByteDance has released on it.

See it for yourself

Start with the output, because that is what earns the usage. These are real Seedance 2.0 generations from Melius, one model across four different jobs: a product beauty shot, a character moment, a cinematic frame, and a scene with a lot of motion. Not a cherry-picked hero reel, just the kind of work people run through it every day.

A quarter of every video, and climbing

Seedance is not something people just experiment with and forget on the platform. It is behind 26.7% of every video generation on Melius, one of only two models that carry the bulk of video work here, and around 4,400 teams have used it. Last week its share sat near 30%. What makes that striking is how fast it happened.

That entire climb ran on Seedance 2.0 alone. Nobody was pushed onto it or was even the model that we automatically routed to. People tried it against everything else we offer, side by side, and kept coming back. That is the useful signal, and it is why a generational update matters more here than a version number usually would.

There is a tell in how they use it. In our data, where we can see the mode, image-to-video beats text-to-video by more than a 2:1 ratio. This is the proof that Seedance is used for production quality videos, not just novelty prompting. People are not prompting Seedance with “cat astronaut” to see what happens. They are bringing a frame they already validated, a product, a character, a locked composition they love, and asking the model to move it and use it as a reference without breaking it.

What AI video generation models struggle at

A clip that looks good for three seconds is a solved problem. Most modern models clear that bar. The hard part, the part that decides whether AI video is a toy or a tool, is everything after those three seconds: holding a face across a shot, keeping a product on-model, running long enough to be usable, carrying its own sound, and not forcing you to throw away a great take because one detail came out wrong.

Read the Seedance 2.5 feature list with that in mind and it stops looking like a spec dump. Every headline item is aimed at exactly those production problems. One caveat before the list: Seedance 2.5 is not public yet, so what follows is ByteDance's announcement, not a benchmark.

• Up to 50 reference inputs, from 12. Four times the old ceiling, spanning images, audio, 3D white models, and style guides. That is what it takes to hold a character or a brand steady across a longer shot instead of watching it drift frame by frame.
• Native audio, in one pass. Audio and video are generated together, in the same latent space, so sound is matched to the action rather than dubbed on afterward. This is the feature people already fake today with a separate audio step.
• Local editing. Fix one region of a clip without regenerating the whole thing, so a single wrong detail no longer costs you the take you liked. At 30 seconds a clip, that is the difference between an edit and a reroll.
• Longer clips. Up to 30 seconds in a single generation, with a beta mode reaching three minutes. Long enough to be a finished shot, not a loop.

The references jump and the audio change are the two most consistently reported, and the two that map most directly onto how people already work. ByteDance also reports 4K capabilities with 10-bit color and roughly 20% better prompt adherence, though the 4K figure is muddier than the rest, since the same event also gave Seedance 2.0 its own native 4K tier. It was announced June 23, 2026, and is to release soon.

Get ready for it

Seedance 2.5 joins the Melius video picker the moment public access opens. Three things set you up before it does.

1. Keep building with Seedance 2.0. It is one of the strongest video models on the platform today, and everything you learn about prompting it carries straight over.
2. Gather your references. If you have been rationing reference frames to fit the old limit, start collecting the full set now. 2.5 gives you room for up to 50.
3. Compare it against the field. Run the same shot through Seedance, Kling, and Veo side by side. Melius allows you to mass produce videos while comparing them to each other, so you get the best chance of getting the video you want.

If 2.5 delivers even most of what is on the sheet, it sharpens every reason people picked it in the first place.

Want to see how your own generations break down by model? Open Melius.

Frequently asked questions