Karta główna porównująca FastFrames Video Generation i Seedance 2.5, lewa karta z tekstem „81 klatek, 5,06 s”, „832x480, bez dźwięku” i „36 ms na klatkę”, prawa karta z tekstem „30 sekund w jednym uruchomieniu”, „30 referencji obrazów i 10 referencji wideo” oraz „rozliczane tokenami”, z logo OrcaRouter w prawym dolnym rogu.
Engineering & Research

FastFrames Video Generation vs Seedance 2.5: trzydzieści sekund kontra osiemdziesiąt jeden klatek

Autor

jinhao.song

Data publikacji

Najnowsze modele · 20Zobacz wszystkie modele →
Benchmarki: Artificial Analysis · aktualizowane codziennie
Powrót do wszystkich wpisów

A thirty-second brief and an eighty-one-frame budget are two different answers to the same question, and the question is not "how fast". Seedance 2.5, ByteDance's long-form entry from 31 July 2026, generates thirty seconds in a single run and extends beyond that across turns, reading up to 30 image references, 10 video references and 10 audio clips as context. FastFrames Video Generation, Apple's 16 November release with Cornell, generates 81 frames at 832×480 and 16 frames per second — 5.06 seconds — in about 2.9 seconds of wall clock on one B200, with no reference inputs of any kind. If you have ever watched a director try to hold a performance across a cut, you already know which of those two numbers is the harder engineering problem and which is the harder scheduling problem. This page is about the second one, because it is the one that decides budgets.

Two metering philosophies, and neither is a price list

Start with how each model bills, because the units do not convert and every comparison that pretends otherwise is arithmetic on a category error.

• Seedance 2.5 — metered in video tokens by ByteDance on the official rate card, published through Volcano Engine Ark in China and BytePlus ModelArk internationally. What a clip costs depends on resolution, duration and actual token usage, so a per-second figure is valid for one resolution and one duration only. Failed takes bill the same as successful ones.

• Third-party hosts for Seedance 2.5 — each adds its own margin and publishes its own per-second or per-clip rate, which is why a dozen pages give a dozen numbers and none of them is the vendor's.

• FastFrames — no meter. You buy the GPU. One B200 for the 1.3B model, eight for the 14B to stay interactive. A generation that fails costs the same seconds as one that succeeds, and both are measured in electricity rather than dollars per output.

• The one independent data point — Artificial Analysis records Dreamina Seedance 2.5 at $34.12 per minute on the text-to-video board and $40.82 on editing, read 8 October 2026. That column is the cost of a minute of 1080p output at the model's default settings, and Seedance 2.5 defaults to the longer, heavier configuration its rate card prices for. Use it for orders of magnitude, not for a budget line.

Which means the honest statement is that one of these models has a price you cannot predict from the outside and the other has no price at all. FastFrames' unpredictability is on your side of the ledger — you know how many GPU-hours you burned — and Seedance 2.5's is on the vendor's, because the token count depends on how the model chooses to spend tokens on your prompt. Between those two, the second is the more dangerous one to plan against.

What the reference budget actually buys

The published reference numbers read like a specification contest and they are really a capability claim.

• Images — Seedance 2.5 accepts up to 30 against FastFrames' zero. There is no image conditioning path in FastFrames at all.

• Video — up to 10 clips as references. This is the one that changes what the model is for: a reference set that includes motion, pacing and camera behaviour drawn from existing footage is a different object from a reference set that only carries appearance.

• Audio — up to 10 clips. Seedance 2.5 generates with sound; FastFrames is silent, and there is no audio path in the released configuration.

• Prompt length — no published cap that matters at this level of specificity for Seedance; FastFrames takes a text prompt and a seed, and its 81-frame window means you are describing one continuous shot rather than a scene.

Thirty seconds with a video reference is a shot you can direct. Eighty-one frames from a prompt is a shot you can sample. Both are legitimate production methods; they are not the same job, and the reference budget is what makes them different jobs rather than different quality tiers.

A two-column scoreboard comparing FastFrames and Seedance 2.5 across six dimensions: single-run length 81 frames against 30 seconds, image references none against up to 30, video references none against up to 10, audio none against generated with up to 10 audio references, resolution 832x480 against 480p and 720p on third-party hosts, and price GPU time you supply against token-metered billing.

Why a five-second window is not the limitation it looks like

Five seconds sounds disqualifying next to thirty, and the architecture is the reason it is less bad than it reads — while also being the reason it cannot be fixed by asking for more frames.

FastFrames is autoregressive: each 81-frame block is conditioned on the one before it, which is why clips chain at all. The authors' contribution is not a faster transformer but the removal of the two components around it that were never the point. A 63.87 million-parameter pixel decoder replaces the VAE decoder and takes decode from about 147 milliseconds a frame to roughly 1. A small refinement network replaces the KV-cache update at 2 milliseconds on the 1.3B model and 6 on the 14B. The transformer itself is frozen in the final stage. On the small model, 271.8 milliseconds a frame becomes 35.8.

Chaining, though, is not extension. There is no keyframe, no start frame, no end frame, no reference image, no clip to continue — so a one-minute sequence built from FastFrames is eleven chained blocks with no mechanism for holding a character's face, a costume or a light direction across the seams. Thirty seconds in one run gives Seedance 2.5 a single context window in which consistency is the model's problem rather than the editor's. That difference is worth more than the sixfold duration number suggests, and it is not something hardware improvements will close.

Where the speed lands you, and where it does not

Convert 36 milliseconds into the units a product person cares about, and the picture is narrower and more interesting than the headline.

• FastFrames-S on one B200 — 36 milliseconds a frame, about 28 requests a second, a 5-second clip in 2.9 seconds. Faster than playback by a factor of about 1.7.

• FastFrames-L on eight B200s — 45 milliseconds a frame, about 22 requests a second, 3.6 seconds for the same clip.

• FastFrames-L on one GPU — 158 milliseconds a frame, which is 12.8 seconds for a 5-second clip. Slower than watching it. The large model is interactive only across a node.

• At 720p — 127 milliseconds on the small model, 116 on the large one across eight GPUs. That is where single-stream interactivity stops.

• On H100 rather than B200 — a 25 to 50 percent penalty by the paper's own estimate, landing the small model somewhere around 45 to 54 milliseconds a frame.

Seedance 2.5 publishes nothing comparable, because it is sold in seconds of output rather than in frames per second of wall clock, and no third party has published a per-frame latency for it that this page could verify. That asymmetry is not a knock; it is a reminder that the two models are instrumented to answer different questions.

What the independent record says, on the one side where it exists

Seedance 2.5 has a board position and FastFrames does not, and both facts need stating plainly.

On the Artificial Analysis video-editing board, read 8 October 2026, Seedance 2.5 sits second at 1,161 Elo over 5,162 votes — editing a clip from a text instruction with the audio preserved. That is downstream of the video-reference capability, and it is the kind of number that means something specific: several thousand people chose its output over another model's output on a task both were asked to perform. It is not a verdict on quality in general, and it does not tell you what the clip cost.

FastFrames has no board entry because it has no host. Apple published weights and no endpoint, so no preference test can serve it to voters, and the model's quality claim rests entirely on its own authors' Table 1: VBench totals of 84.23 for FastFrames-S and 84.44 for FastFrames-L against 84.11 for the Causal Forcing++ baseline at 272 milliseconds a frame, with Wan2.2 14.3B at 84.65. Unreproduced, and on a benchmark that rewards exactly the temporal smoothness autoregressive models are built to produce. The spread across Apple's entire table is about two and a half points while the latency spread across the same models is four orders of magnitude — which is the strongest argument in the paper, and it is an argument about what benchmarks measure rather than about FastFrames' quality.

Costing a real brief

Put an actual job against both and the units collide in a way that is more useful than any table.

• "A thirty-second product film with a locked voice-over" — Seedance 2.5 can do it in one run with multi-turn extension, taking the voice as an audio reference and the product stills as image references. FastFrames cannot hold the audio, cannot take the stills, and would need several chained blocks whose seams you would then have to hide. The job goes to one model on capability grounds before price enters the conversation.

• "A hundred variations of a five-second loop for an interactive display" — FastFrames on a B200 produces each in under three seconds of wall clock, at a marginal cost that is a rounding error once the card is amortised. Seedance 2.5 would bill a hundred generations, several of which would be rejected, against a rate card that does not publish a clean per-second figure to plan with.

• "One hero shot that has to look like the footage we already cut" — Seedance 2.5, because the video reference is the only way either model can be pointed at existing footage. FastFrames has no path from a clip in to a clip out.

• "Prove the interaction concept before we buy hardware" — neither, directly. Seedance 2.5's per-second cost is knowable only in arrears, and FastFrames needs the hardware first. The cheap move is to prototype against a model with a published per-second rate and a real latency figure, then decide.

OrcaRouter does not serve FastFrames — there is no hosted endpoint for Apple's weights anywhere, which is a statement about the release rather than about routing. What one key does reach today is the hosted video layer with published rates: MiniMax H3 at $0.08 per second at 768p and $0.13 at 2K, and the Kling 3.0 family, alongside 200-plus language models on one OpenAI-compatible endpoint at provider list price with 0% markup. Seedance 2.5 is not a route here either — ByteDance sells it through its own platforms and the token card is theirs — and this page is not pretending otherwise. The practical use of that catalogue in this comparison is measurement: get a real latency and cost baseline from a model whose meter you can read, before committing to either a GPU fleet or a token bill you cannot forecast.

The open question, and it is about meters rather than models

If ByteDance published a token-rate example table at more settings, the metered-versus-predictable asymmetry would get much easier to plan around, and a reader could finally compare Seedance 2.5 to a per-second model without doing the vendor's arithmetic for them. If Apple hosted FastFrames at any price, the model would acquire an arena entry and a latency figure somebody else measured — the two things that would let its 36 milliseconds be judged rather than admired.

Neither has happened. Until one does, the way to choose is to notice which question you are actually asking. Thirty seconds in one run answers "can this model hold a scene together". Thirty-six milliseconds a frame answers "can this model keep up with a person". They are both hard, they are both worth paying for, and no single rate card prices both.

The Hugging Face model card for apple/FastFrames-L, showing the title 'FastFrames: Video Generation at Interactive Latency', both checkpoints described as autoregressive two-step video diffusion models generating 480p at interactive latency, the Wan-AI/Wan2.1-T2V-14B base model, arXiv 2611.23226 and the apple-amlr licence tag.The FastFrames project page from Apple and Cornell University, naming the seven authors and stating that FastFrames-S at 1.3B and FastFrames-L at 14B generate 480p at under 50 milliseconds per frame, with tabs for an interactive demo, abstract, quality versus latency, pipeline and comparisons.