The AI Video Prompt Tricks Seedance 2.5 Retired, and the Four That Still Matter

Half of a good prompt was insurance, not direction

Open any ad team's prompt document and you will find the same paragraph pasted at the bottom of every shot. No subtitles. No captions. No music bed. No lettering anywhere in frame. Nobody wrote that as craft. Somebody wrote it the afternoon a finished clip came back with a caption burned across the actor's chin, and kept it because the next clip arrived with a soundtrack no one had asked for.

Prompt folklore is what accumulates when a model has a habit and you have a deadline. Over the past two years ad teams built an entire oral tradition of it: sentences that exist to suppress a behaviour rather than to request one, actor descriptions copy-pasted word for word between variations, punctuation used as routing so a spoken line gets performed instead of read out as stage direction. It worked often enough to feel like skill.

ByteDance's Seedance 2.5 announcement, published on its Seed research blog on 2026-07-31, is the clearest evidence yet that much of it was never skill. It was a patch. Read against the previous version's release note, the new one lands like a list of the exact things ad makers had been writing around, answered inside the model instead. That matters more than any single spec, because it changes what a prompt is for and it tells you which habits in your document are still earning their place.

The folklore that grew around AI video prompts

Three families of workaround dominated, and they all share a shape. The writer is not describing the ad. The writer is defending against the model.

• Suppression: sentences that forbid a behaviour, not sentences that request one.
• Repetition: the same actor description reproduced identically to pin down a face.
• Routing: punctuation and casing used to tell speech, sound and on-screen copy apart.

Each of the three was a rational response to something the model did. That is exactly why each of them was fragile.

The suppression tail

The most common one is a negative list stapled to the end of every prompt: no text, no logos, no watermarks, no subtitles, no on-screen lettering. It exists because the model was building the picture and the words in the same pass. fal's own Seedance 2.0 prompting guide is explicit about that mechanism, saying the model builds the audio and any on-screen text straight from the prompt. Anything the model can author, it can author uninvited, so the suppression tail is a fence rather than a direction.

The actor paragraph nobody wanted to write twice

The second is the copy-pasted actor description: the same forty words about a woman in her late twenties, olive skin, hair tied back, cream linen shirt, reproduced byte for byte across ten variations in the hope that identical input yields an identical face. It does not, reliably, because a description covers a range and a range has edges. We wrote a full production sequence for this in how to keep one AI actor consistent across every ad variation, and the honest summary of that piece is that the fix was never a better paragraph. It was supplying a file instead of a sentence.

Punctuation as routing

The third is the least folkloric of the three, because it was documented from the start. On fal's Seedance 2.0 guide the instruction is plain: put any spoken line in double quotes, and the model lip-syncs it, generates the voice, and times it to the cut. Quotation marks became a channel selector. Ad teams then extended the idea well past what any documentation supported, inventing bracket conventions and casing rules to keep narration, sound design and on-screen copy from bleeding into each other. That extension is where folklore genuinely began, and it is the part worth being careful about even now.

What the model absorbed, in ByteDance's words

Everything in this section is what the announcement states. None of it is a measured result, and none of it carries a published benchmark. Attribution is not a formality here, it is the difference between a fact and an ad.

References instead of adjectives

The headline change is capacity. ByteDance says users can now input up to 30 images, 10 video clips, and 10 audio clips as reference materials in a single pass. Set that against the same company's Seedance 2.0 launch post, which described users simultaneously inputting up to 9 images, 3 video clips, 3 audio clips, plus natural language instructions, and the direction is unmistakable. Fifteen reference files became fifty in one version. The budget for showing the model what you mean more than tripled, while the budget for telling it stayed exactly where it was.

That is the whole argument about actor drift, made in inventory rather than in prose. ByteDance frames the payoff in consistency terms too, saying the model can preserve the appearances and voices of multiple characters while keeping each subject's characteristics stable, even in complex scenarios like multi-character shots or group storytelling. Treat that sentence as a vendor claim. Treat the reference count as the mechanism underneath it.

Subtitles and background music, hedged on purpose

The line ad makers should read twice is this one: the model also minimizes uncontrolled occurrences in subtitles and background music. Minimizes. Not prevents, not eliminates, and with no number attached. That is the announcement conceding, in its own copy, that the failure your suppression tail was built for still happens, just less. We will come back to what that means for your prompt document.

Editing addressed by time, not by re-rolling

The third absorbed workaround is the ugliest one in practice. When a near-perfect thirty-second take had one wrong beat, the only lever most teams had was to generate the whole thing again and hope. ByteDance says Seedance 2.5 offers timestamp-level control for targeted editing of audio and video content, notably improving efficiency and controllability, and that it enhances advanced editing features such as green screen, camera perspective, and reference-based editing. We wrote about why re-rolling a winner is a bad trade in local video editing for ads; the economics there do not change, the tooling does.

Length, which quietly removes a different workaround

The announcement also states that Seedance 2.5 extends single-pass video generation from 15 to 30 seconds. That retires an entire production habit, which is writing a thirty-second spot as six short clips and then spending the afternoon hiding the seams. What Seedance 2.5 is covers the single-take argument in full.

The version note that made the direction obvious

Here is the part most coverage skipped, and it is the reason this story is about causation rather than features.

The maker published the debt first

ByteDance's own Seedance 2.0 launch post does not read like marketing at the end. It states plainly that there is still room for optimization regarding multi-subject consistency, text rendering accuracy, and complex editing effects, and adds that Seedance 2.0 is still far from perfect, with various flaws remaining in its generation results. Three named weaknesses, from the maker, on the maker's own page.

The next release note answers them by name

Now line them up. Multi-subject consistency is answered by the multi-character preservation claim and the tripled reference capacity. Complex editing effects is answered by timestamp-level control and the expanded editing feature list. Text rendering accuracy is the one that is only partially addressed, and it is addressed with a hedge rather than a fix.

Two out of three, in the next version, in the same vocabulary. That is not a coincidence of phrasing. It is a roadmap being executed in public, and it was legible a whole version early to anyone who read the limitations paragraph instead of the highlights.

The rule this gives you

Two reading rules fall out of this, and they cost nothing to apply:

• On the maker's limitations list: engineering debt with a schedule. Your workaround is a bridge, and you should expect to dismantle it rather than refine it.
• Absent from every maker surface: either it is not a model problem or nobody has agreed it is one yet. Your workaround is load-bearing, so document it properly.

The second case is the one worth spending real craft on. Anything in the first case is a habit you are renting.

The same reading rule works on ad platforms, not just model makers. When a platform documents a limitation, treat it as dated rather than permanent: OpenAI's product-feed documentation said for months that catalog campaigns could not use conversion bidding, and it now says the opposite, which is one of several changes landing alongside the pixel default that flips on August 17.

What the route we actually call enforces

Everything above is ByteDance describing ByteDance's surfaces. The announcement is explicit about that scope, saying Seedance 2.5 is rolling out on Jimeng AI, Doubao Pro, and other platforms, with API access coming soon via BytePlus ModelArk. If you reach the model through an API, none of those sentences are a promise about your request.

The contract, field by field

So we checked. On 2026-08-07 we probed the API route Novoads calls for Seedance 2.5 and read the enforced input contract rather than the marketing page. What it accepts:

Input	Enforced ceiling
Reference images	up to 30
Reference videos	up to 10, 30s total
Reference audio	up to 10, 30s total
Clip duration	4 to 30 seconds
Resolution	480p or 720p only

The reference ceilings match the announcement exactly, which is the good outcome and not the guaranteed one. The resolution row is the interesting disagreement: there is no 1080p or 4K tier on this route at all, whatever a spec sheet elsewhere suggests, so anyone planning a high-resolution master off a release note would have been wrong in a way no prompt could rescue.

A vendor number is not a route number

This is a discipline, not a slogan, and we have paid for it recently. In the same week we verified the numbers above, a separate image model on our stack turned out to accept a materially different number of reference files than the figure we had written down from documentation. Nothing was broken. The documentation was simply describing a different surface than the one our requests hit.

The cost of probing is one API call. The cost of assuming is a feature built on a ceiling that does not exist, discovered by a customer. Prefer the call.

Leaderboards carry the same gap in a friendlier costume. A model can sit near the top of a public arena board and still not be reachable from your pipeline, which is exactly the situation we walked through with xAI's Imagine Image 2.0. Rank is not availability, and availability is not your route's input schema.

What the route does not give you

The timestamp-level editing claim is the clearest example of scope mattering. It is a capability the announcement describes on ByteDance's own products. An API route exposes what its input schema exposes, and reference arrays, duration and resolution are not a timeline editor. Until a provider publishes that surface, treat targeted editing as a thing the model can do somewhere, not a thing your pipeline can call.

Which workarounds are retired

Re-describing the actor every time

Retired, and it was always the weakest of the three. The replacement is not a shorter description, it is no description: point the model at reference material and let the file carry the identity. Our breakdown of the reference-to-video input contract covers how those arrays are addressed from inside the prompt, and the tags are positional indices rather than names you invent.

Writing a thirty-second ad as six clips

Retired at the model level by longer single-pass generation. The stitching workflow was never craft either. It was an artefact of a fifteen-second ceiling, and everything teams learned about matching lighting and pacing across cuts was learned to hide a limitation.

Re-rolling the whole take to fix one detail

Being retired, on ByteDance's surfaces first. This one had a real cost attached: every regeneration is a fresh sample, so fixing a wrong label could hand you a worse read of the hook.

The workaround	Why it existed	Status
Actor paragraph	descriptions have range	replaced by references
Six-clip stitching	15-second ceiling	replaced by 30s single pass
Full re-roll to patch	no targeted edit	addressed, vendor surfaces first
Suppression tail	uninvited captions and music	still useful, minimized not fixed
Quoted dialogue	routing speech vs narration	still the documented syntax

Which four still matter

Reference hygiene is the new prompt hygiene

Thirty image slots is not thirty extra wishes. Reference material carries wardrobe, light and room along with the face, so a sloppy set produces a consistent ad you did not want. The craft moved rather than disappeared: it now lives in curating a small, deliberate reference set and reusing it byte for byte, which is a librarian's job more than a writer's.

• Reuse the same files across a variation set, unchanged, or the set drifts with them.
• Keep the reference count deliberate. Contradictory references are worse than fewer references.
• Name what each file is for in the prompt, since the tags are positions in the array rather than labels you invent.

That last point is where teams arriving from image tools get caught. If you are choosing between engines rather than tuning one, our head-to-head on Kling and Seedance sets the two input envelopes side by side.

Minimizes is not eliminates

Keep the suppression line. The announcement's own verb concedes the residual, and there is no published rate under it. A sentence that costs nothing and catches a tail case is a good trade forever, and the day the model genuinely stops authoring uninvited captions, you will be able to delete it in ten seconds.

On-screen text is still a compositing job

Text rendering accuracy is the one item from the 2.0 limitations list that the newer release note does not claim to have solved. If a headline, price or legal line has to be exactly right, render the clip clean and composite the type afterwards. That is not a prompt failure, it is a division of labour, and it will outlive several model versions.

One subject, one action, one camera move

The oldest discipline in the Seedance prompt guide survives everything above, because it is not a workaround. Filmmaking direction is the actual instruction set. Every generation you spend describing what the model can see in a reference is a generation not spent telling it what to do with the camera.

How Novoads solves the prompt-folklore problem

Seedance 2.5 is live in the Novoads model picker, next to Seedance 2.0, Kling v3 Pro, Sora 2, Google Veo 3.1 and Omni Flash, and the route above is the one your generations run through. The practical value of that is not the model badge, it is that you do not have to maintain the folklore yourself: you upload a product photo, write or auto-generate a script, pick an AI actor, and the platform handles the reference plumbing, the model swap per placement and the credit maths per length.

The part worth doing yourself is the test, and it takes two generations:

1. Take one prompt from your document, strip the suppression tail and the actor paragraph, and run it against a real product with your reference set attached.
2. Run the same prompt with the tail restored and nothing else changed.

Compare the two. You will learn more about which half of your folklore is still load-bearing than in a month of release notes, and you will learn it on your own product rather than on a demo reel. If you want the cost side of that comparison first, our Seedance pricing breakdown shows how per-length credit maths works before you run anything. You can start with the $1 trial, which is three days of access, then $49 per month.

The fix landed in the model, not in the wording

The tempting read of the last two years is that ad makers got better at prompting. The evidence says something less flattering and more useful: they got good at describing the shape of a defect, and the defects were then fixed by the people who could actually fix them.

That is worth internalising, because it tells you where to spend attention next. Reading a maker's limitations paragraph now beats reading ten prompt guides, since the paragraph tells you which of your habits has an expiry date. And it sets the correct posture toward every release note, including this one. The claims in Seedance 2.5's announcement are ByteDance's claims, unbenchmarked and scoped to ByteDance's own products until the route you call says otherwise. Verify on your own surface, in your own account, on your own product. A prompt trick that survives that test was never folklore in the first place.