ByteDance set the pace again. At its Volcano Engine FORCE conference the team unveiled Seedance 2.5, and the spec sheet is a lot to take in. In the same wave of news, Seedance 2.0 quietly went 4K — that part’s available right now, and it isn’t cheap — and I got an early look at a Kling research project called OmniDirector that points at where camera control is going next.

A few weeks ago I flagged the 2.5 rumor — early July, name unconfirmed, hold it loosely. The name’s confirmed now, and the July window held up.

Full breakdown’s in the video; here’s the part you can use.

Seedance 2.5: the spec sheet does the talking

As of the announcement, 2.5 isn’t live yet — the current target is early July 2026 — but here’s what’s on the card:

Native 4K, baked into the model rather than bolted on after.
Single shots up to 30 seconds, roughly double 2.0’s current ceiling. The spec card reads as “single shot,” which I’d chalk up to a translation quirk — I can’t imagine those 30 seconds all have to be oners, so I’m assuming multi-shot is in there too.
Up to 50 multimodal references per generation. Fifty. The announcement says that spans images, video, audio, and even 3D assets.
Smart, localized editing — targeting specific parts of a generation for regeneration — shown in a timeline UI that, frankly, has a bit of a Premiere look to it.
And the center square of every model-update bingo card: stronger consistency and control.

Fifty references is the one that turned my head. In theory that’s your entire lookbook per generation. In practice, most shots in the sizzle reel used around four references each — and the open question with any reference system is always how much the model actually respects what you hand it, which can be a bit questionable. Still: 50 is 50.

The localized editing is the other piece I want to see in the wild. They show someone working in what looks like a timeline-based editor later in the video — it’s probably capped footage, but if it’s even half as full-featured as it looks, that’s a meaningful shift from “generate and pray.” We’ll see how deep it actually goes.

No word on pricing yet, and 2.0 was never a budget model — so I’d brace for a new-car tax on 2.5. This is a remarkable amount to pack into a point-five update; it does make you wonder what 3.0 looks like.

Seedance 2.0’s 4K is live now — is the juice worth the squeeze?

The 4K everyone wants from 2.5 is, in a more limited form, already here on 2.0. The lineup now runs from a mini to a fast to a regular to a heftier top-tier flavor — and the real question isn’t the flavor, it’s the cost.

Two caveats up front. First, it’s slow: per usual for a fresh Seedance rollout, my generations sat parked at 99% for hours. I got a couple off anyway — my looping Coda benchmark and a straight image-to-video test — and they both cleared the only bar that mattered here: are they actually 4K? They are. Second, there’s open speculation that this isn’t truly native 4K but a sneaky upscaler doing the work behind the scenes. I can’t confirm that either way — but I can confirm it’s slow, and it costs a lot.

On the cost: there’s always a fog of war between credits and actual dollars, so take the figures as I read them off Dreamina. A 15-second run priced out around 360 credits at 720p; about 885 at 1080p; and a whopping 1,890 at 4K. You can back-trace the math from there.

That’s too much squeeze for me unless you already know the generation is exactly what you want — which, let’s be honest, basically never happens on the first roll. The more cost-efficient path, if you need 4K, is to generate at 1080p and upscale with something like Topaz (see the flamethrower-girl-vs-groundhog-day-girl fight in the video for that test). There are costs there too, so you’re not beating the house entirely — but at least you’re not burning credits hunting for the right shot. If you’re dying to try the native 4K anyway, it’s live now. Just budget the render time.

Kling’s OmniDirector: reference-based camera control, cuts included

Sliding over to Kling. OmniDirector is a research project out of Kuaishou’s KwaiVGI group — the Kling research team — aimed squarely at the camera-control crowd. To be clear up front: this is a research preview, not an announced product. But it’s a strong tell of where they’re headed, and I’d bet we see a feature like it land around Kling 4.0.

It works a little like Kling’s existing Omni reference, except instead of cloning a subject it clones camera movement from a reference clip — and that includes the cuts. One demo actually matches the timing of a cut from its source; another borrows the crane move straight out of Furiosa. There’s weirdness in it — a figure in one shot seems to be mimicking the propeller blades behind her — but that’s exactly the kind of happy accident that makes a tool like this worth experimenting with.

And yes, the dog in the demo is a very good boy. I don’t care if he’s AI or not.

Not available yet, but I’ll let you know the moment there’s real Kling news on it.