How to Use Seedance 2.5 for Ads Without Burning Credits

Seedance 2.5 rewards a timed shot list, not a description. Write your beats at roughly one-second granularity with no gaps, give each reference image exactly one job, put dialogue inside the prompt with in-band markers, and write the literal string `Hard cut to.` between beats you want to land as cuts. Budget three to ten generations per keeper.

Everything else sits downstream of that. It is ByteDance's text-and-image-to-video model with joint audio, up to thirty seconds in one pass, launched 31 July 2026 in the same window as a separate native-4K upgrade to Seedance 2.0, which is where the "2.5 is 4K" story started. Expect thirty to seventy dollars of compute for a finished thirty-second ad at 720p, and more at 1080p, where several providers publish no rate at all.

Is Seedance 2.5 4K? The developer surface says no, the marketing page says yes

No developer route I could find exposes a 4K generation tier for Seedance 2.5. The consumer Dreamina surface advertises and appears to deliver 4K, which is what an export or upscale step after generation would look like, and ByteDance has not published which it is. If a client is buying native 4K delivery, confirm it on the exact surface you are buying, in writing, before you quote. My own parameter reading on 10 September 2026, off the Higgsfield catalogue: 480p, 720p and 1080p for 2.5, 720p by default, no 4K option, against a 4K option still present for 2.0.

That trade looks deliberate: ByteDance bought duration, reference capacity and editing with the resolution tier, a good deal for vertical social and a poor one for broadcast.

Where the myth came from

ByteDance launched 2.5 and separately gave 2.0 a native 4K upgrade, which is how English coverage merged them (Chinese coverage kept them apart). Then its own consumer page for 2.5 went out titled "Official Seedance 2.5: 4K & 30s AI Video Generator with Audio", selling 4K output and watermark-free downloads on the paid tiers, so the press error never got corrected. The research launch post covers single-pass generation, multi-round extension, fifty references and timestamp-level editing without naming a single pixel dimension.

Resellers landed wherever the copy took them. fal puts 2.5 at "480p, 720p" and gives 2.0 "480p, 720p, 1080p, 4K". Morphic describes 1080p as a native, non-upscaled 1920x1080 frame that arrived after launch. Segmind, testing hands-on, wrote that "Seedance 2.5 does not do 4K, and it does not do 1080p" on their surface. Cellcog calls the 1080p-and-above tiers on resellers "provider-side rollouts or super-resolution upscales", Atlas Cloud excepted. Anything above 720p is provider-dependent and moving.

The content farms are what Google ranks today. seedance.tv gives 2.5 "Native 4K, 10-bit color", which it calls "a real jump from the 720p/1080p era", and lists a "Seedance 2 Mini" tier at three credits per second as if it belonged to 2.5, when Mini is a 2.0-family budget tier on every catalogue I have read. MindStudio publishes a correct, hands-on review on the same domain as an article saying the model "adds 4K output resolution". I have asked BytePlus whether the Dreamina 4K path is native or an upscale.

The rest of the spec sheet

From the parameter surface as I read it on 10 September 2026, plus one provider's published reference:

• Duration is an integer from 4 to 30 seconds. One provider's docs default it to -1, letting the model pick its own length and bill on the actual output.
• Frame rate is fixed at 24fps. Output is MP4, H.264 at 480p and 720p, H.265 at 1080p, mono audio, URLs signed for 24 hours, so your pipeline has to pull the file down.
• `generate_audio` defaults to true, which surprises people who then wonder where the voice came from.
• Aspect runs auto, 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, adaptive forced for edit and extension jobs.
• No `negative_prompt` field appeared on any surface I read.
• Seeding is supported, though the same seed gets you the same neighbourhood without reproducing the output.

Where to run it, and what a second really costs

Consumer surfaces (Dreamina, Jimeng, Doubao) sell a credit balance and a UI; developer routes (BytePlus ModelArk, Volcengine Ark, and resellers like fal, Replicate, WaveSpeed and Kie) sell an API and per-second billing. At volume you want the second, because credits price in units you cannot compare across vendors. Per output second, at rates Cellcog verified on 22 August 2026:

The 720p spread runs from $0.20 to $0.473, so the same second costs 2.4 times more depending only on whose endpoint you hit, and picking a reseller on brand familiarity is a tax on everything you generate. Which model to buy is a separate question. Cellcog also says the public ModelArk table had no visible 2.5 row when they checked, so even the "official" rate is derived from the token formula.

That formula matters twice over. Billing is by video tokens from width, height, frame rate and duration, and BytePlus runs two rates, $10.70 per million video tokens without video input and $6.40 with. Nearly every provider in Cellcog's survey switches to a lower per-second rate applied to input plus output duration when you feed a reference video, Atlas Cloud excepted. So a ten-second reference plus a ten-second output can bill you for twenty seconds, which makes reference-to-video dearer than text-to-video at the same clip length, so read your own provider's card before you plan a reference-heavy shoot.

Where can I use Seedance 2.5 for free?

Dreamina and Doubao hand new accounts a starting credit balance, and several resellers give a small standing allowance: Morphic advertises up to 20 credits on a plan it calls forever free. That is enough to watch the model move and nowhere near enough to finish an ad.

Latency, and why you cannot do live client review

Segmind timed 18 completed jobs at 60.1 to 226.7 seconds, median 138.9, and five identical ten-second requests came back between 77.3 and 223.5 seconds, a 2.9x spread driven by queue depth rather than your duration setting. Batch your work and stop promising clients a live session the infrastructure cannot deliver.

Moderation also checks the finished video, so a prompt that sails through submission can still be killed at the end. A request naming film stocks was refused after the generation completed, and reframing it as lighting and lens behaviour got it through. Segmind does not say whether that refusal was billed, and I could not find a provider that documents it either way, so assume you might be charged and keep brands and film stocks out of prompts.

What are the four Seedance 2.5 modes, and when do you use each?

Almost no guide tells you what each mode bills on, which is the part that costs you money.

The `video_edit` row catches people: set duration to eight, feed a twenty-two-second source, and you are billed for twenty-two.

How do I extend a Seedance clip past 30 seconds?

Run `video_extension` against a finished clip with `extension_mode` set forward or backward, and it bills the added duration only, aspect inherited from the source. ByteDance documents multi-round extension, and in my experience each round compounds drift in the face and the product, so two rounds is the realistic limit.

How to use Seedance 2.5 for ads: a shot list with a clock

Veo and Sora reward a well-written descriptive paragraph, which is why people arriving from them write bad Seedance prompts: this model wants a timeline. Runware's docs put it plainly: "Write time ranges at roughly one-second granularity and keep them continuous, with no gaps between windows." Leave a gap and the model improvises filler into it; overstuff a window and you get a dropped beat or a cut you never asked for.

The corollary is the expensive part. fal's prompting guide states it exactly: "The 30-second setting only changes the available duration. It does not add more events to the prompt." Buying thirty seconds for a six-second idea buys twenty-four seconds of drift at full rate, so set duration to the length of the events you wrote.

Most guides converge on the same six-part grammar: subject, action or event, scene and environment, visual style, camera movement or cut, and audio. Segmind publishes it as ByteDance's six-part formula, though that is their reading of the grammar ByteDance shipped rather than a quotation from it. Front-load hard, because the model locks subject and core action from the opening words, and practitioners put the weight in the first twenty to thirty. That matches my experience, and nobody has published a controlled test of it.

Two camera rules hold up everywhere. One principal move per clip, because pan plus orbit plus crane inside thirty seconds gives muddy motion with no readable geometry. And frame position instead of adjectives, since "the red player stays in the left third of frame, hoop visible on the right" is a constraint the model can satisfy while "dynamic tracking" is a vibe it has to guess at. Keep camera and action in separate clauses, since merging them flattens both, and write causation as timed blocks (contact, then movement, then sound, then reaction).

How do I bind reference images to the right job?

Bind each reference explicitly, in prose, and forbid it from doing anything else. fal's guide gives the template: "@Image1 controls only [identity, product, wardrobe, environment, or another invariant]. Do not copy [pose, background, lighting, text, or camera angle] from @Image1." Most people skip the forbidding half, which is the half that stops your studio backdrop leaking into a kitchen scene.

Three rules make it survive a real shoot:

1. Confirm the upload order of your media array first, because reference numbering follows it. Get that wrong and nothing errors: the prompt runs, binds the wrong image to the wrong job, and bills you full rate. That is the bug I have personally wasted the most money on.
2. Declare a priority wherever two references could fight over a property. "Keep the product geometry higher priority than any shape visible in the motion clip" costs eleven words and saves a reroll.
3. Attach two to four references for most shots. The documented ceiling of fifty is a trap, since conflicting references controlling one property is a top cause of wasted spend, and every hands-on guide converges on a small pack, under eight image subjects even for complex shots. Multi-angle sets of one product (front, side, detail) beat a single hero photo, because they tell the model how it is built.

Tag syntax belongs to the wrapper. Square brackets (`[Image1]`) appear in provider-facing docs for the ByteDance routes; the at-sign (`@Image1`) is what my pipelines and fal's worked examples use. Sources also disagree on whether the tags are case-sensitive, with at least one provider writing them lowercase throughout, so match your provider's docs and test on one cheap 480p generation.

How do I write dialogue so Seedance lip-syncs it?

Dialogue synthesises and lip-syncs at no extra charge, which removes the separate TTS pass most UGC pipelines were built around. The syntax uses in-band markers, documented consistently across Morphic and several prompt guides:

Plain double quotes also trigger voicing and sync on most surfaces. Put language, accent and delivery before the line, because the model reads the instruction and then performs the text: "Spoken language: American English. She says it fast and a little annoyed: {I told you this would happen.}"

The one line that fixes talking-head ads

One line buried in a worked example in fal's guide does more for talking-head ads than anything else in this piece:

Without it the model animates the mouth through silence, a more obvious tell than the hands or the skin.

How many spoken words fit in a clip?

From my own pipeline across roughly forty finished vertical ads rather than any published source, so calibrate it against your own voice talent:

A thirty-second ad is therefore around sixty to seventy spoken words, which forces the discipline that makes these ads work: every plot event that can be shown gets shown, because visual beats cost zero words. The wince, the package landing on the counter and the product held up to the light buy you story for nothing, while echoes and repeats carry no information.

Two more script rules from the same pipeline. Emotion has to live in the words, because the model under-renders flat prose: stretch the vowel on the peak word, allow a volume spike or two per line, break one sentence at the peak. And never write an engineered dramatic pause, because pauses bloat the line and break the render.

How do I get real hard cuts instead of a morph?

It takes two changes, applied together.

The first is a literal string, out of the UGC workflow I run rather than any provider's documentation: put `Hard cut to.` at the end of every cut description except the last, and Seedance treats it as an edit instruction. Runware never mentions hard cuts and fal's worked examples steer the model away from them, so there is no published test of this, I have not run a controlled one, and the effect I see is confounded with the delta rule below. Take it as pipeline doctrine and spend twenty 480p generations proving it to yourself.

The second is the delta rule. A boundary only snaps into a crisp cut when the two adjacent beats are visually far apart, and low-delta neighbours blend however many times you write "hard cut". Every adjacent pair has to differ on three axes at once:

1. POV alternates every slot: selfie, then locked-off static, then selfie, never two consecutive slots in the same point of view.
2. Distance band rotates every slot. Tight (close-up, macro), mid (medium, medium-close), wide (three-quarter, waist-up, full-body, product-extended). Adjacent slots come from different bands, and across eight slots each band appears at least twice.
3. A different physical action every slot, so the same hand-and-product configuration never runs twice.

Shifting the angle or micro-location so the background changes is a bonus, and the strongest cut-forcer of the lot. A default eight-slot cadence: selfie-mid, static-wide, static-macro, selfie-tight, static-mid, static-macro, static-wide, selfie-tight. State framing distance in capitals every slot, because the model responds to `MEDIUM CLOSE-UP` and ignores "she is fairly close to camera".

The storyboard sheet

The board trick comes out of a production workflow I run and I have not seen it written up anywhere public. Instead of generating eight clips and fighting drift between them, build one wide 21:9 sheet holding eight vertical 9:16 panels in a row, feed that sheet as a reference in `omni_reference`, and the model renders one continuous clip with eight internal hard cuts. Panel K becomes cut K, and the person, wardrobe, lighting and product hold because all eight panels were painted together in one image pass.

Two limits. The prompt is still the primary signal and the board secondary, so a sparse prompt gets the board copied frame-for-frame and looks posed. And the practical ceiling is around fifteen seconds per clip, so longer ads mean concatenated boards (4 to 15 seconds is one, 16 to 19 is two balanced ones, 20 to 30 is fifteen plus the remainder), appending the previous board as the final image media from board two onward to carry continuity. Assembly is a stream copy, because every join is already a hard cut:

`ffmpeg -f concat -safe 0 -i clips.txt -c copy final.mp4`

How do I keep a character consistent across shots?

Seedance has no character memory across a cut, so consistency is something you re-assert every beat. Repeat one fixed physical description verbatim in every cut, with no synonyms, and restate the full invariant list after any occlusion: same face, same hair, same coat, same boots, same walking speed and direction. Lock a second character to a fixed left or right screen position and pick a contrasting archetype, since two similar figures swap features. Give any character who matters a reference image, because a described face is not anchored the way a referenced one is. The face drift people report is almost always the verbatim repetition or the post-occlusion restatement being skipped.

Why does my product change size or shape between shots?

Because Seedance 2.5 ships without a product LoRA, an IP-Adapter or a fine-tune, so fidelity comes down to reference binding plus prose, which makes it a writing problem with no technical solution.

Enumerate invariants as nouns, since every part you name is a part the model will not redesign. That is why fal's worked example spends its words on "matte cobalt-blue shell, black rubber grip ring, circular copper button, and clear lower chamber" while "sleek modern design" fails to survive a camera move.

Then lock the angle: the product shows only the side visible in your reference, keeps it in every slot, and must not rotate, spin, flip or reveal unseen faces, switching between reference angles only across a hard cut. Left to itself the model invents back labels, side panels and internal components, all of it gibberish, and it enlarges products so the label reads, so forbid that and give size relative to the hand ("palm-sized, fits entirely in one hand, roughly 15cm tall"). Specify component states up front, lid on or off, cap twisted or seated, or a viewer notices a cap that re-closes itself and cannot say why the clip feels wrong. The model also forgets across an occlusion, so when the product or the person passes behind something, repeat the full invariant list for the re-emergence.

Lock the ending too. The last two to three seconds want a stable medium close-up, product upright and front-facing, label unobstructed, no camera movement, no rotation, no cut, no fade to black, which hands you a clean pack-shot frame for type. Nobody claims Seedance reproduces small legal copy or tight kerning at 720p, so generate the motion plate and composite real brand assets over it; ending a prompt with "ready for a price and logo overlay" is how a professional briefs a plate. One staging trick, called a workaround rather than a feature by its users: include a clay render of the product alongside the photographs, so the model reads dimensions unambiguously and the bottle stops resizing when the camera orbits.

Other failure modes, and the prompt-side fix for each

Morphing cuts, silent mouths moving, resizing products and drifting faces have their own sections above. These are the rest, with the fix I write for each; where no source is named, it comes out of my own pipeline.

Does Seedance 2.5 have a negative prompt field?

No. There is no `negative_prompt` parameter, which is awkward given how much of the SERP tells you to use one while another chunk warns that naming an unwanted element makes it more likely to appear. What works is a terminal negative block in prose at the end of the prompt, the placement Runware's docs describe, where directives like "No subtitles" or "No BGM" are interpreted reliably. Concrete beats generic: "No product deformation, no changes to the logo, label text, bottle shape or packaging proportions, no duplicated bottles" does work that "avoid artifacts" never will.

Can I use a real person's face as a reference in Seedance 2.5?

Not directly. BytePlus says it restricts making videos from images or videos that contain real faces, framed as mitigation against impersonation, and the restriction runs across the BytePlus, CapCut and Volcengine surfaces. Two routes are sanctioned: real-person verification and likeness authorisation through the ModelArk console for someone whose permission you hold, or ByteDance's library of over ten thousand virtual human assets across a range of ages and backgrounds.

That policy is why the whole AI-UGC genre generates a synthetic actor in an image model first and feeds that face into Seedance, which is the shape the policy leaves available. Sanctioned is a long way from risk-free: a generated face that resembles a real person still creates exposure under right-of-publicity and digital-replica statutes, which attach to likeness rather than to how the pixels were made. If you want the person to be real, get the authorisation.

How do I make ten ad variants from one winning video?

Performance work is where Seedance 2.5 earns its keep, through `video_edit`. It takes one source ad between four and thirty seconds and produces N independently edited versions, holding the source's motion, performance, camera, cuts, lighting, pacing, aspect ratio and exact final duration, so you keep a proven control constant and vary one variable, which is the closest thing to a real A/B test in generative video.

What follows is one pipeline's configuration, the Higgsfield ad-multiplier workflow, so treat the parameter names as that wrapper's contract rather than ByteDance's API, the `@ImageN` convention and the 3,900-character cap included. On the raw model API the values differ: Runware's editing docs require duration set to `"auto"`, reject an integer outright and refuse width and height alongside a source video. Check your provider before you copy any of it.

It renders silent, which turns out to be the trick: set `generate_audio: false`, extract the source's audio and remux it afterwards, so the winning voiceover and music survive byte-identical across every variant. Duration is `ceil(source_duration)`, aspect on auto.

The text-preservation block is mandatory, once per prompt, verbatim, or your captions come back re-rendered as approximations:

For a person swap, state that the original must never appear in any frame, extended through cuts, entrances, exits, occlusions, motion blur, transitions, reflections and shadows, or the original comes back in a shop window while you stare at the new actor. A person reference covers the complete look by default, face, hair, skin tone and texture, build, grooming, clothing, footwear, headwear, eyewear and jewellery, so if you want the new person in the old clothes, attach a garment image and say so. Ordered lists zip by position, so ten swaps produce ten videos. Retry a failed position at most once, bind every replacement with `from @ImageN` rather than describing the identity as coming from the source, and use the long-form numbered-operation template for any identity swap.

Run this only on footage you own or have a release for, and make sure the release covers identity replacement and derivative edits, because a creator who agreed to appear in one ad did not agree to be swapped out of ten. Feeding a competitor's winning ad into this pipeline is not what the mode is for.

Making it not look AI, which is where the money is

The best evidence points somewhere uncomfortable: the click-through penalty attaches to looking AI, whatever the image is. Exner, Hartmann, Netzer and Zhang, across TUM, Columbia and Harvard Business School, ran a quasi-experimental sibling-ad design on Taboola: 4,633 matched ads, around 369 million impressions and 2.5 million clicks, differing only in whether the image was AI-generated. Raw comparison put AI images at 0.76% CTR against 0.65% for human images, and with experiment fixed effects that gap collapses to parity, which the authors read as AI ads performing comparably at far lower production cost. Human raters then scored 1,751 images blind for perceived artificiality (460 AI, 1,291 human, five raters each):

The "not AI" rows pool images rated definitely human-made, likely human-made or not sure; the paper's own term is that the image disguises its origin. AI images that disguised their origin beat every human image in the set, and human images that read as AI came last. The paper names intense colour saturation as a cue signalling AI generation to consumers, and medium to large faces as reading human. It covers static images on one platform's inventory, with advertisers self-selecting into the AI tool, though the direction is evidenced well enough to build a process against.

Counter-evidence points the other way hard. An NYU Stern finding carried in version 2 of the IAB's AI Transparency and Disclosure Framework (18 August 2026) is that telling consumers an ad was made with generative AI cut click-through by 31.5%, and Gartner's survey of 1,539 US consumers, fielded October 2025 and released 16 March 2026, found half preferring brands that avoid generative AI in consumer-facing content. Put those beside the sibling-ad study and the honest conclusion narrows: removing AI cues helps where no label applies, and where a label does apply you are optimising the wrong variable.

What the model stamps on your output, and what you owe in disclosure

BytePlus says it is implementing C2PA Content Credentials on top of its existing visible watermark capability, and the Dreamina surface carries a visible AI label too, with watermark-free downloads sold as a paid-tier feature. Behaviour differs by surface, so verify on the route you buy before you promise a client a clean plate. Stripping the label or the provenance metadata may breach your provider's terms and, in some jurisdictions, the law, and this article does not advise it.

You may also owe a disclosure of your own. Article 50 of the EU AI Act has applied since 2 August 2026: providers must mark synthetic output in a machine-readable, detectable format, and deployers who generate or manipulate image, audio or video deepfakes must disclose it as artificially generated, clearly and at the latest at first exposure. Breaches sit in the Article 99 tier reaching 15 million euro or 3% of worldwide annual turnover, whichever is higher. Meta, TikTok and Google run their own labelling policies on top of that, and China's rules are stricter again. I am not your lawyer: get advice before you scale.

The de-slop pass

I run every storyboard through an image-to-image pass before video, at the frame stage where a grade can no longer reach. It holds framing, composition, poses, subjects and product where they are and changes only micro-realism, identically in every panel: in goes pore-level skin, real material detail, even daytime light, faint sensor noise and a flat phone-photo look; out go waxy skin, beauty-filter smoothing, over-saturation, HDR bloom, oversharpening, the teal-orange grade and shallow depth of field. Keep each face's shape, width and proportions one to one too, because image models love to slim faces and a slimmed face reads synthetic immediately. Half the things a stock "cinematic" prompt asks for are the exact cues that make an ad read as AI, so the prettier preset is the one that loses you money. The longer version of this recipe, and the evidence behind it, sits in the piece on running this as an ad business.

Frozen-frame QA before anything ships

Step through evenly spaced frames, every product close-up, and two or three mid-word frames.

1. Exactly one hero product, no clones anywhere in frame.
2. At most two hands per person, counting mirrors and partial hands at the frame edge.
3. Absent features stay absent, and cap, button and prop states stay consistent across cuts.
4. Labels are not gibberish, not mirrored, and have not drifted into a different real brand, the failure that gets a takedown rather than a bad ad.
5. Product scale matches the holding hand.
6. No doubled lip edges, no face drift, no baked-in text or subtitles.
7. Visible AI label present or absent, and provenance metadata intact.

Remedies differ by cause: a staging failure means correcting the prompt and rerunning that clip only, lip artefacts mean cutting spoken words first, baked text gets one rerun before you remove it in post. Never bake captions in generation; burn them post-render from a word-level transcript, where you control font, timing, spelling and line breaks.

What does a finished 30-second Seedance 2.5 ad cost?

The credit-price pages answer a question nobody has. What you want is the price of a deliverable:

cost per finished second = (raw cost per generated second ÷ keeper rate) ÷ edit retention rate

Keeper rates decide it. Poppify's field aggregation puts simple static shots at one to two generations per usable result, mixed-complexity work at three to five, and six to ten for hands, lip-sync, walking, multiple subjects and hand-object interaction, which describes every UGC ad ever made. So, for a thirty-second vertical UGC ad built as two fifteen-second boards at five attempts per board:

That excludes reference-video input seconds, moderation refusals, the image passes and every hour of human labour. Most shops shoot this as a 1080p job, the row where several providers publish no rate at all.

A vendor with an interest in the answer lands in the same window: Higgsfield's pricing post puts a finished thirty-second video at $30.60 to $69.12 across five platforms, assuming four shots at three attempts each, and notes that four of the five entry plans run out of credits before the job finishes.

The two published yields both come with a catch. PJ Accetturo's Kalshi NBA Finals spot is reported at 300 to 400 generations for 15 usable clips, roughly four percent at about $2,000, though I could only trace that count to secondary coverage, it was made on Veo 3, and we have numbers for that ad because it worked, which is the wrong direction of selection for anyone budgeting. invideo's own pages disagree: its FAQ puts roughly 25 percent of clips in the final cut against a cost breakdown documenting 108 images and 103 videos generated to land 49 used clips, which it calls roughly 50 percent utilisation. Take the range, and expect your own number to depend more on how tightly you brief than on which model you bought.

The one saving that beats every per-second discount

Approve a still before you spend a single generation. Get sign-off on a keyframe, then go image-to-video from it, which moves the revision cycle to the cheapest artefact in the process. It roughly halves generations per keeper in my own work across a couple of dozen briefs, so treat that as one operator's figure.

Secondary savings, in descending order: edit a region or a clip instead of rerolling everything, draft at 480p and approve at 720p (480p runs a bit under half the 720p rate), test at the shortest practical duration, change one variable per retry.

Compute is not your biggest cost anyway. In the one itemised sixty-second breakdown I could find, editing labour was 44 percent of a $229.80 stack against 31 percent for video compute, and that source is vendor-adjacent, so trust the ranking over the numbers. Writing and directing sits at zero in every vendor comparison I have read, which is convenient for the vendor and expensive for you, and the economics of this as an ad business turn on that line more than on any per-second rate.

Should I use Seedance 2.5 or Seedance 2.0 for ads?

2.5 also costs meaningfully more. MindStudio's review puts 720p at about 23 cents per second against roughly 15 for 2.0, calling the quality gain modest and mostly in motion consistency and prompt following. That matches my experience: what you buy with 2.5 is duration, references and the edit modes.

Is Seedance 2.5 better than Veo 3.1 for ads?

For vertical UGC, usually yes: Seedance gives you thirty seconds in one pass, fifty references and `video_edit`, where Veo 3.1 works in much shorter clips with a handful of reference ingredients and costs more per second on every rate card I have compared. Veo still wins on 4K delivery, on physical realism through fast motion, and on Google's published indemnity for generative services, which matters to some clients more than anything else. The full model-by-model comparison goes through the trade in detail.

What the leaderboards do and do not tell you

Be careful with any "tops the leaderboard" claim about this model, in either direction. Every figure below was read on 10 September 2026.

Seedance 2.5 does not appear on Artificial Analysis's text-to-video board at all; Seedance 2.0 sits at rank 5 with an Elo of 1222. On arena.ai's text-to-video board it does appear, at rank 6 on 1482±12 across 48 models and 668,045 votes, just ahead of 2.0 on 1479±8. On arena.ai's video-edit board, 2.5 at 720p sits second on 1410±26 behind Wan 3.0 on 1414±26, from 429 votes in a category holding ten models (nine when the operator first publicised the result), so first and second are four points apart inside error bars of twenty-six. Anyone still quoting the "#1 in Video Edit" headline is quoting a snapshot that has moved.

The rank-6 placing does not change the argument, because the sample is portrait-heavy and the error bars overlap. Wireflow puts more than sixty percent of arena test cases as leaning toward portrait and talking-head scenarios, so a model tuned for close-ups wins a board made of close-ups, which tells you little about a product orbit or a walking shot. No consumer leaderboard measures inter-shot consistency or on-screen text rendering, the two things deciding whether an ad is shippable, and a September 2026 paper reports inter-shot consistency far below intra-shot (arXiv:2609.06373). VBench-2.0 has no entry for 2.5 either, and its authors note that many aspects of superficial faithfulness are approaching saturation (arXiv:2503.21755).

Best fit and worst fit for Seedance footage

It is strong on macro product beauty where physics matter and humans do not (liquids, pours, steam, powders, textiles), on single-move hero shots ending on a clean pack-shot frame, on vertical UGC talking-heads as one continuous take, on on-body products where the body gives natural scale, and on stylised work where an AI look is no defect. For liquids, write contact, then flow, then pooling, then separation, and say where the motion settles. Physics stays weak across every frontier model on published benchmarking, so "strong on liquids" means strong relative to peers rather than reliable. Where I stop reaching for it: fast motion, multi-character interaction, small legible text, high-consideration purchases where a viewer will study the product, and anything needing a specific real person.