Seedance 2.5 is Here: 30 Seconds, One Take, Sound Included

ByteDance shipped Seedance 2.5 today, and it's already live in the Video Studio. Text to video and image to video, both of them, right now.

I'll be honest about why I'm excited, and it isn't the benchmark chart. It's that the thing I kept fighting with in every other video model mostly went away. You know the fight: you get five seconds, it looks great, and then you need a second shot, so you generate again and the character's jacket is a different color and the room has moved and the music restarts. You spend the afternoon stitching. Seedance 2.5 does the whole scene in one pass.

Thirty seconds, and it's actually one scene

Up to 30 seconds in a single generation, with synced audio. That number matters less than what happens inside it. 2.5 will hold multiple connected shots in one clip and let a story move through it, a setup, something developing, a turn, a resolution. Cuts, camera moves, the same face across all of them, sound that stays glued to the picture the whole way.

Compare that to what you were doing last week: generate 5s, generate another 5s, pray, open an editor. Now you write the scene and it comes back as a scene. I ran a "detective walks into the rain, stops, turns around, delivers one line" brief as a first test and got all four beats in one file with the line landing on the right lips. That was not the normal experience two weeks ago.

Durations in the picker go 5, 6, 8, 10, 12, 15, 20, 30. Aspect ratios include 21:9 now, which is a small thing that makes a big difference if you're cutting anything that wants to look like film.

Write a brief, not a prompt

This is the part people will miss, so I'm putting it in a heading. Seedance 2.5's prompt window is huge, we allow up to 10,000 characters on it, and it rewards you for using them.

Stop writing "cinematic shot of a woman in a cafe, 8k, film grain." Write the thing you would hand a crew. Who's in it, what they're wearing, shot one, shot two, where the camera is, what the light is doing, what we hear, what she says. The long production-brief style prompts are what this model is built for, and the difference between a lazy one-liner and an actual brief on 2.5 is bigger than the difference between most models.

Up to 30 reference images

Image to video takes up to 30 reference images per generation, against 9 on the 2.0 endpoints. Character from one, wardrobe from another, the location, the prop, the lighting mood, a style plate. If you've been building a look across a project and re-explaining it in words every time, this is the fix. Output follows your source image, so no aspect ratio picker in that mode, it just matches what you gave it.

ByteDance's own announcement also covers video and audio references, timestamp level editing, green screen background swaps, and extending a clip across multiple rounds into several minutes. Those are real and they're great, and right now the studio exposes the image path. We're working through the rest. I'd rather tell you what's actually wired up than let you find out by hunting for a button that isn't there.

It's the new default when you don't want to think

If you use the auto pick and just want the best thing available, that now points at Seedance 2.5 for both text and image. If you don't want to think about models at all, you don't have to, you'll get this one.

2.0 is still there and still very good, and it's cheaper per second. For a quick 5 second beauty shot it's often the sensible call. The price in credits shows before you render either way, so you can see the tradeoff instead of guessing (1,000 credits = $1).

The part that's ours, not ByteDance's

Everything above is the model. Here's the bit I care about: you run it here without your work being logged against your name. Generation goes through our anonymity layer, encrypted on your device, identity split from content by Oblivious HTTP relays, decrypted only inside attested secure enclaves. No single party holds both who you are and what you asked for. The full architecture is here.

Nobody should have to hand over their creative direction, their client's unannounced campaign, or a personal project they're not ready to show anyone, just to use a good model. That's the whole reason this product exists.

Go make something

It's live now, free to start, no card.

Try Seedance 2.5 text to video →

Try Seedance 2.5 image to video →

If you make something good with the 30 second window, I genuinely want to see it. This is the first model where "just write the scene" is real advice, and I don't think any of us know yet what that unlocks.

Seedance is developed by ByteDance. OpenGradient is not affiliated with ByteDance.