Hero card comparing FastFrames Video Generation and Vidu Q4 Preview, left card reading '832x480, silent', '36 ms per frame' and 'text prompt only', right card reading 'up to 4K', '15 image refs, 3 audio refs' and '3 to 16 seconds', with the OrcaRouter logo bottom-right.
Guides & Insights

FastFrames Video Generation vs Vidu Q4 Preview: A 4K Preview Against a 480p Loop

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Vidu Q4 Preview is the odd one in this series, because the word in its name is doing real work: it is a preview build, released 7 October 2026 by Shengshu Technology, and Shengshu's own material frames it as the version that gathers the feedback shaping the final model. FastFrames Video Generation is also not a product, but for the opposite reason — it is finished research published as weights on 16 November, with no endpoint, no price and no roadmap to either. Put them side by side and you get two models that are both unavailable in the sense a procurement officer cares about, on opposite sides of the same gap: Vidu Q4 Preview has a shipping API and a rate card and will be superseded; FastFrames has no API at all and will not be. What follows is about which of those two absences costs you more.

The rate card is the first genuinely comparable thing

Vidu Q4 Preview is one of the few models in this field that publishes a per-second rate by resolution tier, which makes its arithmetic legible in a way most of its competitors' is not.

• 540p — 9 credits per second, about $0.045 at Shengshu's published standard credit rate of $0.005.

• 720p — 19 credits per second, about $0.095.

• 1080p — 24 credits per second, about $0.12.

• 2K — 38 credits per second, about $0.19.

• 4K — 78 credits per second, about $0.39, or roughly $0.72 for a sixteen-second clip at list rate before discounts.

• FastFrames — nothing. No rate, no host, no queue. The cost is one B200 for 2.9 seconds of wall clock on the small model, or eight B200s for 3.6 seconds on the large one, per 5-second clip.

Two things about Vidu's curve are worth stating because they are easy to misread. The $0.014 per second figure that accompanied the launch announcement is the 540p tier at a discounted credit rate, not a general rate. And the platform is running limited-time package discounts of up to 30% on credit bundles, which move every tier down together rather than changing the shape of the curve. Plan against the list-rate curve; treat the discount as a temporary offset that may not be there when your project ships.

The comparison that falls out is not "one is cheaper". Vidu Q4 Preview has a floor — a 540p blocking tier priced for exactly what it is used for, proving a composition before paying for a full-resolution render — and FastFrames has no tiers at all, because it has no billing surface. If your work involves iterating a shot at low resolution before committing, Vidu's 540p tier is a real product feature and FastFrames' fixed 480p is not the same thing, even though the numbers are close.

Two models built for different ends of the same brief

The capability lists barely overlap and that is the most useful fact in this comparison.

• Resolution — Vidu Q4 Preview generates natively at 540p, 720p, 1080p, 2K and 4K, with 10-bit colour at the top two tiers. FastFrames is 832×480 with no setting and no upscale path.

• Duration — Vidu Q4 Preview runs 3 to 16 seconds on Image-to-Video and 1 to 16 seconds on Reference-to-Video. FastFrames is 81 frames at 16 fps, 5.06 seconds, extended only by chaining.

• Inputs — Vidu Q4 Preview documents two endpoints, /ent/v2/img2video and /ent/v2/reference2video, taking up to 15 images across characters, scenes and style plus up to 3 mp3 audio references of 3 to 12 seconds for voice cloning. FastFrames takes a text prompt and a seed. Not a smaller reference budget — no reference budget.

• Mode coverage — Vidu Q4 Preview documents no text-to-video path and no video-editing path. FastFrames is text-to-video only. Between the two of them, every mode the other one lacks is missing from both.

• Audio — Vidu Q4 Preview generates audio and video together, with an audio parameter in the image-to-video endpoint outputting dialogue and sound effects plus reference-voice consistency. FastFrames is silent.

• Latency — FastFrames publishes 36 milliseconds a frame on one B200, 45 on eight for the 14B, and 158 milliseconds on a single GPU for the large model. Vidu Q4 Preview publishes no per-frame latency figure that this page could verify.

Read those six lines together and the two models are almost complementary. Vidu Q4 Preview is the one that holds a cast, generates sound, and delivers 4K; FastFrames is the one that produces frames faster than you can watch them, from text, at 480p, silently. Neither substitutes for the other on any of the six.

A two-column scoreboard comparing FastFrames and Vidu Q4 Preview across six dimensions: resolution 832x480 against up to 4K, clip length 81 frames against 3 to 16 seconds, image references none against up to 15, audio references none against up to 3, audio none against generated dialogue and sound, and price GPU time you supply against 9 to 78 credits per second.

The board position, and why it is a caution rather than a verdict

Vidu Q4 Preview has an independent score and it deserves to be read carefully rather than quoted.

On the image-to-video board at Artificial Analysis, read 8 October 2026, it sits third at 1,179 Elo over 5,543 votes. First is MiniMax H3 Max at 1,195, second is MiniMax H3 at 1,181. Sixteen points separate first from third, on intervals of roughly ±10 and ±11. That is a statistical tie dressed as a podium, and any page that opens by declaring a winner on that board is reading noise. What the votes do establish is that Vidu Q4 Preview's output is competitive with the top of the field on a task thousands of people have now judged.

FastFrames has no such record and cannot acquire one while it is unhosted. Its quality claim is its authors' own VBench numbers — 84.23 for the 1.3B checkpoint and 84.44 for the 14B, against 84.11 for the Causal Forcing++ baseline at 272 milliseconds a frame and 84.65 for Wan2.2 14.3B at 117,734 milliseconds. Nobody independent has reproduced them, and VBench rewards the temporal smoothness autoregressive models are good at. The table's entire quality spread is about two and a half points; its latency spread is four orders of magnitude. That is a fair argument that the benchmark has stopped discriminating, and an unfair way to claim a model is as good as its neighbours.

One more caution about Vidu's number specifically: the votes are from October 2026 and were cast against a field that has since changed. A model entering the arena later, or a preview build being replaced by a general release, re-sorts the board without any of the listed models improving or degrading. Read the date with the figure, always.

What "preview" costs you, and what "research release" costs you

Both labels are honest and both have commercial consequences worth spelling out.

A preview build is a licensing and continuity risk. Vidu Q4 Preview's pricing and feature availability are, by Shengshu's own framing, things that settle at the full release — which means the rate card you budget against today may not be the one you are invoiced against in six months, and a feature you build on may change shape. The compensating advantage is that the preview exists: there is an API, there is documentation, there are two named endpoints, there is a support path, and there is a stated direction of travel.

A research release is a different risk. FastFrames' weights carry Apple's AMLR terms, the LingBot-World v2 demo components in the same repository are CC BY-NC-SA 4.0 — non-commercial — and the rest is under Apple's Sample Code Licence. There is no endpoint, no SLA, no support and no deprecation policy, because there is no service. What you get in exchange is permanence and inspectability: the weights will not be withdrawn, the paper will not be rewritten, and Apple published every training-stage checkpoint for the large model — the causal teacher, the consistency-distilled stage, the DMD-distilled stage, and the final generator with its pixel decoder and KV refinement network — which makes the release reproducible in a way almost nothing else in this field is.

The hardware requirement is the dividing line for most teams. FastFrames needs Linux x86-64, an NVIDIA GPU and CUDA 12.8, then uv sync and a generation script. Vidu Q4 Preview needs an account and an API key. That sentence decides the comparison for more organisations than any Elo difference will.

Which one, by brief

Take Vidu Q4 Preview when the job is consistency of cast and voice, when native 4K generation at 10-bit colour is a delivery requirement, when the input is a reference set or a still rather than a prompt, or when a cheap 540p blocking pass before a full-resolution commit is how your process actually works. Accept that it is a preview, that the pricing may settle differently, and that it documents no text-to-video or editing path.

Take FastFrames when the loop is the deliverable — a live preview, a streaming avatar, a controllable world model — and when 480p silence from a text prompt is the right shape for the product rather than a compromise. Accept the licence review, the GPU fleet, the absent reference inputs and the fact that nothing in this release is aimed at a 1080p-with-sound deliverable.

OrcaRouter does not route Vidu Q4 Preview — there is no Vidu identifier in the catalogue — and it does not host FastFrames, which has no endpoint for anyone to route. What one key does carry is the hosted video layer with published per-second 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 pricing passed through with 0% markup and automatic failover between endpoints. For a team weighing a preview build against a research release, neither of which is a safe production bet today, that catalogue is where to get a stable per-second baseline before either decision.

What would settle it

Two events and no others. A full Vidu Q4 release that holds its board position at the same per-second curve would turn the preview into a production commitment and remove the continuity risk; a release that adds a text-to-video endpoint would put it on the same boards as everything else and make the comparison direct. On the other side, Apple hosting FastFrames at any price — or anybody else hosting it — would give the model an arena entry and an independent latency measurement, which is the only thing that would let a reader judge 36 milliseconds against anything rather than admire it in isolation.

Neither has happened as of this writing. Which leaves the decision where it usually is: on the deliverable, on the licence, and on whether you would rather rent a preview or own a paper.

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.