Hero title card for a comparison of FastFrames Video Generation and HiDream O1 Video 1.0, showing two facing cards labelled 'FastFrames — 36 ms per frame, 480p, silent, open weights' and 'HiDream O1 Video 1.0 — Elo 1,175, 1080p with audio, $5.80 per minute', with the OrcaRouter logo bottom-right.
Guides & Insights

FastFrames Video Generation vs HiDream O1 Video 1.0: Latency, Meet Narrative

Author

jinhao.song

Date Published

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

Two releases, ten weeks apart, built on opposite beliefs about what a video model is for. HiDream O1 Video 1.0 started from the assumption that a clip is a short film: it plans a narrative arc, generates five to twenty seconds at 1080p with synchronised audio, and treats physics as a constraint rather than an aesthetic. FastFrames Video Generation started from the assumption that a frame is an event that must arrive before the next one is needed, and it optimised the distance between frames down to 36 milliseconds — at 480p, with no sound, on weights Apple published on 16 November 2026 alongside a paper dated the 17th. Neither is a worse version of the other. They are answers to different questions, and picking one tells everyone which question you were asking.

One is measured by a leaderboard, the other by a stopwatch

The asymmetry in the evidence is the most important thing on this page. HiDream O1 Video 1.0 has been ranked by a third party; FastFrames has not, and could not be in its current form, because a 480p text-to-video model with no audio does not slot onto an image-to-video board that assumes sound.

• HiDream O1 Video 1.0 — sixth on Artificial Analysis's image-to-video board with the audio filter on, Elo 1,175 ±10 across 3,231 samples, at $5.80 per minute, in the 10 October 2026 capture. Artificial Analysis records its release as 28 August 2026, the board entry as 10 September, and HiDream.ai's own announcement as 17 September.

• FastFrames-S and FastFrames-L — no board entry, no independent evaluation. The 84.23 and 84.44 VBench totals are the authors' own, from Table 1 of the paper, on the same prompts and seeds as every baseline in that table.

• The claim that connects them — nothing measured. VBench totals and Elo are not convertible, and no one has run both on the same prompts.

So the honest way to compare quality is to refuse to. What can be compared is capability, latency, price and access, and on three of those four the answer is not close.

A two-column scoreboard comparing FastFrames and HiDream O1 Video 1.0 across six dimensions: latency 36 ms per frame against not published, resolution 832x480 against 1080p, duration about 5-second blocks against 5 to 20 seconds planned, audio none against synchronised, conditioning text only against text plus one reference image, and price GPU time only against $5.80 per minute.

What HiDream O1 Video 1.0 gives you for $5.80 a minute

It is the more ambitious system by a wide margin. HiDream's launch material describes three stages — a planning pass, joint audio-video generation, then an alignment stage — with diffusion reinforcement learning post-training against a multimodal reward model. In practice that surfaces as a video model that decides how long a clip should be rather than being told: duration is narrative-planned within the 5 to 20 second window, and the audio is generated with the picture rather than laid on afterwards.

• Resolution and length — 1080p, 5 to 20 seconds, duration planned rather than selected.

• Audio — generated jointly with the video and synchronised to it, not a separate pass.

• Conditioning — text, image and video together, with exactly one reference image required on the hosted API.

• Physics — treated as a generation constraint, which is the claim behind its above-median Elo.

• Access — a hosted MaaS endpoint with an async task and poll cycle, at $5.80 per minute of output, which is about $0.097 per second.

• Open weights — none.

The API is narrower than the launch material, and that is the detail that catches teams out. The public endpoint accepts exactly one reference image, given as a public URL or Base64 without the data-URL prefix, between 40 KB and 20 MB. The prompt is optional and defaults to empty. There is no reference-video field, no resolution selector — aspect ratio is inherited from the reference — and no audio control. The only knob is a force_10s boolean. And there is a client trap worth knowing before you write the polling loop: the top-level result.status reaching 1 means every subtask has finished, not that they succeeded. The actual outcomes are in sub_task_results[].task_status, where 1 is success, 3 is a generation failure and 4 is a safety refusal. A naive poll loop will mark failed generations as done.

What FastFrames gives you for the price of a GPU

Less product, more machine. FastFrames is an autoregressive video diffusion model distilled from Wan2.1 — the 1.3B checkpoint from Wan2.1-T2V-1.3B, the 14B from Wan2.1-T2V-14B — using a recipe called Causal Forcing++. Its contribution is not the transformer but what surrounds it: the VAE decoder is replaced by a 63.87 million-parameter pixel decoder that emits 4×16×16 RGB values per token across three causal ConvNeXt blocks and a linear head, cutting decode from roughly 147 milliseconds to about 1, and the KV cache update becomes a small refinement network — shared stem, four 8-head attention blocks, one residual repair head per DiT layer — costing 2 to 6 milliseconds. The DiT itself is frozen at the final stage.

• Output — 832×480, 81 frames, 16 fps. About 5.06 seconds per generation, chainable autoregressively.

• Latency — 36 ms per frame for the small model on one B200, 45 ms for the large model on eight; 158 ms for the large model on one GPU, and 127/116 ms at 720p.

• Conditioning — text only. No image, no video, no keyframes, no camera control.

• Audio — none.

• Access — weights on Hugging Face under Apple's AMLR licence, code on GitHub, Linux x86-64 with CUDA 12.8. The bundled LingBot-World v2 demo is CC BY-NC-SA 4.0, which bars commercial use of that part.

• Price — no hosted rate anywhere. You supply the hardware.

The quality claim, for what it is worth as an unverified vendor number: 84.23 VBench total for the small model and 84.44 for the large, against 84.11 for the Causal Forcing++ baseline at 272 ms per frame and 84.65 for Wan2.2 14.3B at 117,734 ms. If those hold, a model running at a fraction of a millisecond per frame more than the baseline matches a model running twenty-six hundred times slower. That is the claim to test, not to quote.

The Hugging Face model card for apple/FastFrames-L, showing the title 'FastFrames: Video Generation at Interactive Latency', both released 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 three-audience decision

Divide the buyers by what they are making and the choice stops being a comparison.

If you are producing finished video

HiDream O1 Video 1.0 is the only one of the two that produces a finished video. 1080p, sound, a duration chosen to fit the content, priced per second so you can forecast. At $0.097 per second a twenty-second clip is $1.94, which is a normal line item on a production budget. FastFrames cannot enter this conversation at any volume, because it cannot deliver a file anyone would watch.

If you are building an interactive system

The reverse is total. HiDream's async task-and-poll cycle is structurally wrong for anything that has to respond within a frame — you submit, you wait, you fetch. FastFrames at 36 ms per frame on one B200 produces faster than 16 fps playback, which is the threshold at which a human stops noticing that a machine is generating. If your product is a live preview, an avatar, or a steerable scene, there is no hosted model in this article that does the job, and Apple's release is the most credible open route to it — subject to the licence review, which for the demo components means non-commercial only.

If you are evaluating whether to build either

Then the cheapest experiment is against a hosted model, and OrcaRouter is one key for the hosted side: MiniMax H3, the closest hosted neighbour to HiDream's ambitions, sits alongside Kling 3.0, Kling 3 Turbo and Kling Video O1 in the same catalogue, at provider list pricing passed through with 0% markup so a vendor rate change lands on your bill the day it happens. Neither HiDream O1 Video 1.0 nor FastFrames is served there — we route neither, and this article is not a claim that we do. What the key buys is a way to measure your own latency tolerance before you spend on the hardware FastFrames needs, because that measurement is the only thing that decides whether 36 milliseconds is worth the 480p.

What would settle this

One experiment: run both on the same twenty prompts and put the outputs in front of people who did not know which model made which. FastFrames will lose on every prompt where sound or detail matters, and it may win on the ones where motion coherence is what the eye catches, because autoregressive generation is good at exactly that. Until someone runs it, the defensible position is that HiDream O1 Video 1.0 is the better video model and FastFrames is the better latency result, and those are two different products being sold — one per second, one per GPU-hour.

The FastFrames project page from Apple and Cornell University, naming the seven authors, stating that FastFrames-S at 1.3B and FastFrames-L at 14B are autoregressive video diffusion models generating 480p at under 50 milliseconds per frame, with tabs for an interactive demo, abstract, quality-latency plots, pipeline, text-to-video samples and comparisons.