Two weeks ago I was stitching together a 12-second Kling clip and a 15-second Seedance clip for a client deliverable, praying the lighting matched at the cut point. It mostly didn't. I spent forty minutes in DaVinci color-matching two AI-generated shots so a jacket wouldn't shift from navy to black mid-scene. That's most of what the job actually looks like these days, which is probably why the Seedance 2.5 announcement out of Beijing three weeks ago got my attention in a way most model drops don't anymore.
ByteDance showed a model that generates 30 seconds in a single continuous pass. No stitching, no seam. If you've never had to hide a stitch point in a client cut at 2am, this might sound like a spec bump. It isn't. It's closer to the line between a toy that makes AI clips and a tool you can actually build a production around.
What ByteDance actually changed with Seedance 2.5
Most of what gets called "long-form AI video" right now is several short clips glued together, and the glue shows. Faces drift a few percent between cuts. Lighting jumps. Motion stutters right at the seam because the model had no memory of what came before. Seedance 2.5 generates the full 30 seconds as one pass, so it holds context the whole way through instead of guessing fresh at every new clip.
There's a second number in this announcement I think is underrated: 50 multimodal reference assets in a single generation, up from 12 in Seedance 2.0. Images, video, and audio all feeding one generation together. On a client job where I'm matching a brand's exact product shots, a locked-down actor likeness, and a specific music cue, that turns three separate fights into one input stack.
None of us can touch it yet, though. It's closed enterprise beta, China-first through Doubao and Volcano Engine, with international access coming later through BytePlus ModelArk and Dreamina. So this is a get-ready story, not a go-use-it story. ByteDance also quietly folded a native 4K upgrade into the existing Seedance 2.0, alongside Seedream 5.0 Pro for images and Seed-Audio 1.0. Which suggests they're building a full stack rather than shipping one flashy model.
Kling's $2.8 billion, and the Sora-shaped hole behind it
Alibaba joined a $2.8 billion round into Kuaishou's Kling AI on July 2, valuing the company at roughly $15 billion pre-money, with BlueFive Capital, Baidu, and Tencent all in the cap table. Alibaba, Baidu, and Tencent don't usually write checks into the same anything. They spend most of their time competing with each other. This time they agree that video generation is the next real platform, not a demo.
This lands right after Kling 3.0 Turbo and Omni shipped on June 17 with 4K editing, a new Omni One engine, and clips that cap at 10 seconds natively but stretch to about 3 minutes via extend. I ran Kling 3.0 on a summit recap project this month, and the multi-shot storytelling mode. Up to 6 connected shots with physics-accurate motion. Gave me the best consistency I've gotten out of a publicly available model. It's not at Seedance 2.5's level, but it's shipping, and shipping beats closed beta.
Both of these stories are really one story: Sora is gone. OpenAI killed the consumer app on April 26, and the API sunsets September 24. The vacuum that left is what Kling's money and Seedance's tech are rushing to fill. If you built a pipeline around Sora eighteen months ago, you're migrating now. Practically speaking, that means testing Kling 3.0 Omni for anything under 3 minutes today, and keeping a Seedance 2.5 enterprise-access request sitting in your inbox for whenever it opens up.
An AI actress, a real film, and an industry that isn't having it
Tilly Norwood, the AI performer built by Particle6 Productions under Eline Van der Velden, is set to star in a comedy-drama called "Misaligned," announced July 6. It's a hybrid shoot: real directors, real writers, real editors, working alongside AI specialists and a lead character that doesn't exist. SAG-AFTRA's response was not subtle. The union called it built on "stolen performances" that devalue human artistry, and stated plainly that Tilly Norwood "is not an actor."
I don't think this is a simple villain story, but I don't think SAG-AFTRA is wrong to be angry either. Somebody's training data made that face and that voice, and nobody's fully accounting for whose performances got folded into it.
Set that against "Dreams of Violets" premiering at Tribeca, the first full-length live-action AI-generated film accepted at a major festival. A 75-minute docudrama built around the January Tehran protests. One project uses AI to manufacture a marketable star. The other uses it to reconstruct a story that couldn't otherwise be filmed. Both are built on the same underlying tools, aimed at pretty different ends: one toward replacing a job, the other toward reaching footage nobody could have shot.
A workflow trick worth stealing this week
If you're generating anything past 8 seconds in Kling or Seedance right now, stop writing one long freeform prompt. Chunk it. Break the scene into 2-3 second beats and treat each one like a line in a shot list: camera movement, subject action, framing, in that order. "Slow push in, subject turns head left, rim light holds" reads completely differently to these models than a loose paragraph of description. It's already the standard trick carrying over from Seedance 2.0 workflows into 2.5, and it's a big part of why Kling's 6-shot storytelling mode holds together instead of falling apart by shot four.
Worth watching from a different angle: political campaigns are already using this same tech, Spencer Pratt's LA mayoral "Batman" spot being the obvious example. Tools built for client car commercials are now writing campaign ads, and nobody's really asked the consent question out loud yet. Between that and Tilly Norwood, the harder questions about this technology look like they'll reach most working creators well before they reach a film festival jury.