Seedance 2.5's Draft Mode Makes It Cheap to Change Your Mind

Just before China's National Day holiday, ByteDance added a small option to Seedance 2.5. Users can now generate a fast 480p sample first and wait to commit to a 1080p final render until they like what they see. It is the kind of feature that sounds like a settings toggle. It is actually a statement about where the real cost in AI video lives.

The cost of a usable clip in AI video is not the cost of one render. It is the cost of every render you threw away to get there. Ask anyone who has spent an afternoon generating takes and discarding them: the expensive part is the search, not the final print. Draft mode attacks that search directly. Exploring becomes cheap. Committing stays deliberate.

Three companies, three answers to the same bill

Seedance's draft mode makes the most sense next to what two competitors did in the same stretch. Together they read like three different answers to one question: how do you make AI video affordable enough to iterate on?

ByteDance's answer is to make the search cheap. Generate a low-resolution draft, run through ten ideas in the time and cost of a couple of finals, and only pay the full price for the take you actually keep. Kling's answer, following the Kling 4.0 release, is to make the search shorter. Improve control with keyframes and multimodal references so that fewer attempts are wasted and the first or second take is closer to what you wanted. Vidu's answer, with its Q4 Preview priced from 0.09 yuan per second, is to make the search irrelevant by pushing the per-second cost so low that people stop counting attempts at all.

None of the three is obviously right. They reflect different guesses about what creators want. Some creators would rather have control and fewer surprises. Others would rather have a low enough price that control matters less. Draft mode assumes something in between: that creators want to think in broad strokes first and refine later, the way a painter works on a rough sketch before committing to detail.

What all three share is a focus on the same underlying number. In video generation, the metric that decides whether a tool is usable is not the price of a render. It is the price of a usable shot, which includes the cost of everything you generated and discarded along the way. A model that is twice as expensive per render but produces a usable shot in half as many tries can come out ahead. Draft mode, keyframe control and rock-bottom per-second pricing are three ways to attack that number, and the winner will be the approach that best matches how a given creator actually works.

Why draft-first fits how people actually work

The draft-first approach matches a workflow that predates AI by a long way. Storyboards, animatics, rough cuts and previz all exist because making the expensive version before the idea is settled is a waste. A 480p sample is the AI-video equivalent of an animatic. It is not meant to look final. It is meant to answer one question: is this the shot I want?

That question usually has less to do with resolution than people expect. A shot either works or it does not, and you can usually tell from a rough version. The camera move feels wrong, the framing is off, the timing drags. None of those problems need a 1080p render to reveal themselves. Draft mode lets you catch them before you pay for polish.

There is a second benefit that is easy to miss. Low-resolution drafts are faster to produce, which keeps the creative loop tight. The longer you wait between an idea and seeing it, the more detached you become from the decision. A fast rough draft keeps you in the flow of making choices, which tends to produce better results than agonizing over a small number of expensive renders.

What Seedance 2.5 already does

Draft mode is not the only thing in this release. Seedance 2.5 extended single-generation length to 30 seconds, accepts up to 50 multimodal reference materials, and added local video editing, so you can change part of a clip without regenerating the whole thing. It also natively supports more than ten languages. Put together, these features point at the same goal as the draft mode: reduce the number of times you have to start over.

That goal matters because of how people use these models now. More than a novelty, video generation has become a tool for testing. Advertisers test several openings for the same product. Short-drama teams test character relationships and camera rhythm before committing to a scene. A draft mode that makes testing cheap, combined with a reference system that keeps a character consistent across takes, turns the model from a slot machine into something closer to a rehearsal space.

Reference consistency is the quieter half of that combination. Generating one good shot is not hard anymore. Generating ten shots of the same person, in the same clothes, in the same light, is still where projects stall. Anything that makes consistency cheaper to achieve compounds with anything that makes raw generation cheaper to attempt, and Seedance 2.5 ships both in the same release.

The honest caveats

Two things can go wrong with a draft-first approach. The first is fidelity. If the 1080p final does not match the 480p draft closely, the cheap exploration loses its point, because you cannot trust what you are choosing. That gap is where the feature lives or dies, and it is worth testing on real prompts before building a workflow around it.

The second is that drafts can become their own time sink. When exploration is cheap, it is easy to explore endlessly and never commit. At some point the cost moves from money to attention, and a creator who generates a hundred drafts is no better off than one who generated two expensive finals, unless the hundred drafts led somewhere. Cheap iteration is a tool, not a substitute for taste.

Still, the direction is right. The industry spent the past year racing on peak quality, and peak quality is approaching the point where most viewers cannot tell the difference. The next competition is about how cheaply and how quickly a creator can arrive at the shot they want. Draft mode is a small feature that answers that question in a specific way, and it is a better sign of maturity than another benchmark record.

It is also the kind of feature that tends to spread. Once one model makes cheap exploration the norm, users start expecting it everywhere, and competitors have to answer. That is how a toggle becomes a standard. In a year, generating a rough version before committing to a final may look as obvious as proofreading a draft before sending it, and the models that made it obvious will be the ones people remember.