ByteDance released Seedance 2.5 on July 31 under the line "one-take creation, flexible referencing." It went to Jimeng Web and Doubao Pro first, then landed on BytePlus ModelArk on August 6, making the API available in supported markets globally.
Not in the US, which matters if your client's team sits in New York. From Bangkok it works, and that's the reason I'm writing about it instead of bookmarking it for later.
AI video has been usable for a while. What's been missing is a version you can drop into a client project without rebuilding the project around it.
Source @bytedance
What actually changed
Thirty seconds in one generation. Seedance 2.0 stopped at 15. Doubling that sounds boring until you try to build a spot out of it. Fifteen seconds is a moment. Thirty holds an opening, a middle and an end, which is another way of saying it holds an idea.
You can extend the result too. ByteDance's own model page says twice; its launch material describes multi-round extension into longer sequences. Either way, the interesting part isn't just duration. It's whether the model can carry the same visual and audio language through the extension without making the edit obvious.
Up to 50 references in one generation. Thirty images, ten video clips and ten audio files, alongside the prompt.
ByteDance also says Seedance 2.5 reads reference video for things like framing, pacing and cinematic language rather than treating it purely as motion to copy. That's the difference between "move like this" and "shoot like this." If it holds up on real briefs, this may be the most useful part of the release.
Audio comes out with the picture. Dialogue, effects, ambience and music can be generated in the same pass, and BytePlus says the model supports more than ten languages natively.
That doesn't mean a proper mix, VO session or localisation workflow suddenly disappears. But for concepting, social content and early client review, having picture and sound arrive together changes the speed of the process.
And you can edit instead of starting over. Timestamp-level prompting to change a particular moment. Localised edits instead of replacing the whole frame. Green screen, camera-perspective changes, reference-based editing and white-model control for blocking shots before committing to a finished render.
This is the feature that decides whether AI video becomes genuinely useful in production. Generating something impressive once is easy to demo. Changing second 14 without destroying seconds 1 through 13 is the harder problem.
BytePlus has also published pricing now: pay-as-you-go is listed at $6.40 per million tokens when video is included as an input and $10.70 per million without it. Those numbers are useful for procurement. They're less useful for production until we know what one approved shot costs after retries.
And there are still reasons not to confuse a launch reel with a production test. ByteDance itself says it wants to keep improving the model's understanding of real-world physics. The demos look good. That's not the same thing as knowing how it behaves on the seventeenth revision of a real brief.
The bottleneck moved
We've been generating AI video on real client work for a while. Coffee films, UGC-style product demos, a tourism hero video, social content for a construction machinery brand, gemstone macro for a jewellery client. It's all on our experiments page.
What that work taught us is that the model was never really the hard part. Continuity was.
Same bottle. Same light. Same face. Same material. Same brand world across eight clips. That's where the hours went. Most of those projects were won or lost on reference wrangling, not prompting.
Fifty references and a one-pass 30-second output go straight at that problem. But the bigger shift isn't that one video gets easier. It's that a family of videos becomes possible.
A client almost never needs one video. They need a website hero, three social edits, six verticals, two language versions and a seasonal cut that still looks like the same brand. Building all of that from scratch every time is part of why video costs what it costs.
Building it from one reference pack, where the product photography, packaging, palette, approved camera language, previous footage and audio identity are inputs rather than instructions, is a system. Systems are the thing an agency can actually sell.
Three things this changes in practice
Concepts move before the budget does
Pitching a 30-second animatic with voice, music and pacing is a different meeting from pitching a moodboard. Clients decide better when they can see the thing. We already do this with stills. Video changes the room.
It doesn't mean the animatic becomes the final campaign. It means you can test three directions before committing the production budget to one of them. That is a much better place to discover that everybody secretly preferred option B.
Client assets start earning
Most clients are sitting on far more useful material than they realise: product photography, packaging files, old campaign footage, 3D models, brand films, sound design, approved voice work. Historically, a lot of that material becomes an archive. With reference-heavy generation, it can become input.
A jewellery brand can hand over product images, a macro reference clip, a material reference and a sound direction, and we can test three treatments before anyone books a studio. A hospitality brand can give us its existing footage, interiors, photography and approved visual language and use that to explore a new campaign without asking a model to invent what the brand looks like from a paragraph.
That distinction matters.
Revisions stop being rebuilds
This is the commercial one. If "change second 14" means regenerating the whole clip and losing the parts the client already approved, you either eat the cost or pad the quote. Neither is a great system.
If it means editing one moment while leaving the rest alone, the maths starts to work. That's less exciting than another cinematic demo reel, but for actual client work it matters more.
The part nobody puts in the launch post
The model doesn't arrive without rules. BytePlus adds C2PA Content Credentials to generated output, uses protection mechanisms around intellectual property, and restricts video generation from images or videos containing real faces.
Verified clients can go through real-person verification and likeness authorisation for specific people who have given permission. Everyone else has access to a library of more than 10,000 virtual human assets.
Read that as an agency and it's basically a scope document. You can't quietly drop a real ambassador into an AI spot because it's convenient. Generated media can carry provenance information. And platform permission is not the same thing as rights clearance. Music, trademarks, likenesses and client-owned material still need the same questions they needed before the model existed.
That review doesn't disappear because the frame was generated.
Where I wouldn't use it
Anywhere the product has to be exactly right. Watch faces. Fine typography. Packaging detail. Jewellery geometry. Anything where being 95% correct is still wrong.
Brand-critical type goes in during post. Generated type has improved enormously, but I still don't want a campaign depending on whether a model remembers the kerning on a bottle label.
Anywhere you need a real person the audience already knows. Founders, staff, ambassadors. We shoot those. There are authorised workflows for likeness now, but "the platform lets us" and "this is the right production decision" are two different questions.
And I wouldn't put it on a hard deadline with no fallback. Not yet. I don't have retry-rate numbers from our own Seedance 2.5 briefs. Until I do, it doesn't go on the critical path.
What we're doing next
Running it through real briefs instead of curated demos. Three questions matter to me.
How many attempts does it take to get one usable 30-second clip? Does a 50-reference brand kit genuinely hold product identity and visual language across a sequence? And does the generated audio survive a client listening on good speakers rather than a phone?
I don't know the answers yet. And anyone pretending they know exactly how this changes commercial production three weeks after launch is guessing.
None of this removes strategy, craft or oversight. It moves where they pay off. When the first direction takes an afternoon instead of a week, the advantage doesn't automatically go to whoever can generate the most video. It goes to whoever can direct the system, keep a brand world consistent, know which idea out of six is actually worth making, and turn the output into something finished.
That part still looks a lot like creative work.
We'll publish what we find the same way we publish the rest of our experiments: outputs visible, failures included.
Restons en contact.
Suivez Asia Media Studio pour des conseils en matière de design et des mises à jour sur Facebook, Instagram et LinkedIn.
