All aboard the Seedance 2.5 pre-release hype train. I’ve got a batch of new samples, some fresh info on when the model actually ships, and an interesting beta feature it looks like it’ll launch with. (Quick aside: the Bilibili crowd translated one of my videos into Chinese, and the comment section over there is arguing about cost and which model already beat Seedance — so, exactly like every comment section everywhere.)
If you missed the announcement itself, that’s the companion piece to this one: Seedance 2.5 Is Official covers the spec sheet, including that up-to-50-references headline. This post is the first look at what the model actually puts on screen.
Full breakdown’s in the video; here’s the part you can use.
Real samples, not the fakes going around
Everything here is sourced from the Volcano Engine Ark page, so it’s legit. Worth flagging because there are a number of fake “Seedance 2.5” clips making the rounds that are really just two Seedance generations stitched together. These aren’t those.
Sample one: eight image references, one continuous shot
The first one runs on eight image references and a very detailed prompt, and the adherence to that prompt is pretty on point as the generation plays out. The thing that impressed me is how it changes styles via the reference images as our guy walks from room to room.
Not perfect, though. In the New Year party scene there’s some morphing and identity loss as he crosses the party — especially under occlusion, where he goes soft and blobby passing in front of another guy. A door also vanishes; can’t tell you if that was intentional. And there’s a missed loop opportunity — he could’ve walked back into the door and reappeared at the start. That said, it’s a busy shot with a lot of particle-y effects going on, and I can’t discount compression on the web version.
Right around the 15-second mark things get a little soft. I’m not saying that’s a last-frame/first-frame trick — I actually think it’s probably generating natively. But that’s something I’ll be testing once we get our hands on the model. And since this is a oner, there’s a decent chance multi-shot dodges the softness entirely.
Sample two: thirteen images plus three video references
Honest take up front — I didn’t love this output. The transition effects don’t do it for me, there are some weird dissolves, and when it cuts, it cuts unrhythmically. I don’t think it’s that great.
What made me sit up was the reference load. This one is packing 13 image references plus three video references. The prompt calls out tracking and orbit, and that shows up in the output. I didn’t Where’s-Waldo every image, but Flamethrower Girl’s long-lost cousin Butcher Guy is in there, along with a mom carting a baby, and the locations all match their image references. The audio it’s referencing is actually the fireworks — not music.
I do wonder if that fireworks audio was generated in SeedAudio 1.0, which we looked at in the last video. In that video I speculated SeedAudio might be baked into Seedance 2.5; I’m starting to lean away from that. I think they’ll end up being two separate models that just happen to work really well together.
One reference is plenty
After all that context overload, there’s a simplified example — one image reference and a long prompt. The takeaway: you can throw the phone book of references at 2.5, but you’ll get a lot more stability if you don’t throw all your cards down at once.
Motion, texture, and whole-shot references
Multi-video references point to some genuinely interesting techniques. One example leans more toward the advertising / motion-graphic side, and what’s notable is that the reference videos aren’t just motion references — they’re texture references, getting a certain gooeyness across to the model so it lands what you’re after. And it looks like you can insert reference material as a whole shot or segment, just wholesale. That part’s fascinating.
The 50-reference question
One of the mic-drop moments from the Volcano presentation was ByteDance flexing up to 50 reference images in a single generation. I tend to read this as expanded context memory — the room got bigger, so you can put more stuff in it. But the room still has limits. So it’ll be on us to figure out the balance: how much do 50 images weigh versus, say, five video references? That’s also the fun part — where you experiment, get results you weren’t expecting, and those turn into techniques down the road.
Multilingual, expanded
Multilingual support has been expanded for both audio and text. One example is a row of people each saying “thank you” in a different language — and each character is image-referenced, with 2.5 doing a good job capturing each character’s essence from a one-shot image. There’s also multilingual support within a single generation. Where I think this comes in handy is localization, especially combined with the editing features.
Video-to-video editing: the one that’s actually about control
Here’s the standout. A fantasy warrior is hanging out in a forest — she’s lost her bow and arrow, but that’s not stopping her from practicing. Through video-to-video editing, she can cast a spectral arrow, and everything else — characters, environment, camera, timing — stays locked in place. That’s the whole pitch: change one element, keep the rest.
(I talked to the character afterward. She wanted it on record that the sound was the bow string. We all saw her at crafty absolutely destroying the cheese plate.)
Fighting dudes — and no celebrity swaps
It wouldn’t be a Seedance video without some fighting dudes. The output’s a little mushy in my opinion, but that’s the kind of thing we’ll all play with and figure out how to sharpen on release. What I can guarantee you 100% is that you’re not going to be able to swap these guys out for Brad Pitt or Tom Cruise. Seedance has learned its lesson there.
The cinematic samples
Rounding out, a couple on the cinematic side. One in the advertising category looks pretty good, especially for text-to-video — textures hold up well. I don’t know if it was originally sourced as a 4K output that we’re now seeing downscaled off the website, but it reads clean. Then a more narrative cinematic sample, with subtitles once again provided by Claude.
The Dreamina leak
Here’s where it gets interesting. The Dreamina Seedance 2.5 page has gone up — kind of. I had to take this one through its paces, because what I think happened is it’s a temp / work-in-progress page. There’s a play button that doesn’t do anything. The user-testimonial section is a dead giveaway — sorry, Olivia Brown, James Wilson, and Sophia Taylor, but you don’t exist. And the “explore more AI models” section lists Sora 2, Veo, and Happy Oyster and Happy Horse, which is a little odd for an official Dreamina page.
But I backtracked all of it, and it is the real Dreamina page. My best guess is an intern hit publish by accident — and that intern is in for a rough day. Under the key features, it’s all the stuff we just covered: reference-to-video, 50 multimodal inputs, video editing, and multilingual creation.
The 180-second beta
The bombshell is buried under a “generate longer videos with stable continuity” section, which mentions a beta mode that extends generations up to 180 seconds. The copy reads like it was copy-pasted from official docs rather than randomly generated — so, potentially three-minute generations in a beta mode.
Is it six 30-second generations stitched together? Maybe. But calling it a beta does seem to hint it might be native. If so, is it going to be perfect? Probably not — that’s what the beta label is for. Will it cost an ungodly amount of money? That’s 100% a yes. But if it gets there natively, that’s a massive barrier-smasher for AI video.
Release window, plus a Fast and a Mini
On timing: we’ve been getting more chatter, and it looks like a global release in about two weeks. I don’t expect the pure chaos of the 2.0 launch — I think this one goes pretty smoothly. And per Emily, a reliable Seedance tracker, two things worth holding onto: with 2.5 around the corner, expect a wave of scam sites, so be careful and be wary; and there’s also going to be a Seedance 2.5 Fast and a Seedance 2.5 Mini.
So that’s it for today. The Seedance 2.5 watch officially begins.