FastFrames L: what Apple’s video model is, and what it is not

FastFrames L (FastFrames-L) is Apple's 14B-parameter text-to-video model, not a frame-interpolation tool, that makes 480p video at interactive speed under a research-only licence; this site currently delivers images only, not video, but its photo editor, a different model, edits one photo in about 37 seconds for 10 credits. Free with sign-up: 10 credits = 1 image, then from $0.13 per image.

Your edited photo appears here (this site’s editor)
Studio background edit made on this site: the same woman on a plain light-grey studio backdrop
Input photo (AI-generated, fictional): a smiling woman with curly brown hair in a green sweater, in a parkYour photo
Type your own change, or leave this empty and pick a one-tap edit below.0/400

Style changes (cartoon, 3D, anime) are not reliable yet. If the result looks unchanged, try again or describe the change more specifically.

One-tap edits

  • Not FastFrames, not affiliated with Apple
  • Images only: this site does not make video
  • Free with sign-up: 10 credits = 1 image
parameters of FastFrames-L (README); S has 1.3B
14B
video size from the README’s example: 81 frames at 16 fps
832 × 480
what the licence file in the repo allows
Research only
the three final-generator files (our sum of the file list)
30.6 GB

What FastFrames L is

From the Hugging Face card and the README

  • Apple released FastFrames-S (1.3B) and FastFrames-L (14B). The README calls both autoregressive, two-step video diffusion models that generate 480p video at interactive latency. The card’s pipeline tag is text-to-video and its base model is Wan2.1-T2V-14B. Neither the card nor the README describes frame interpolation.Source: Hugging Face card (apple/FastFrames-L) ·
  • How L was made, per the README: Wan2.1-14B was first distilled into a two-step causal model following the Causal Forcing++ recipe, then a pixel decoder and a KV refinement network were trained for it. The repo ships the causal teacher, the consistency-distilled model, the DMD-distilled model, and the final generator with its decoder and refiner.Source: GitHub README ·
  • What goes in and out, per the README’s example: a text prompt, a seed and an output path go in, and an 832 × 480 video of 81 frames at 16 fps comes out, which is about 5 seconds by our arithmetic. The README names no image input.Source: GitHub README ·
  • Requirements the README states: Linux x86-64, an NVIDIA GPU and a CUDA 12.8-compatible driver, then uv sync. Checkpoints and third-party weights come from a separate checkpoint guide that we did not open.Source: GitHub README ·
  • Dates: the card gives no release date. The Hugging Face repo was created on 2026-11-16 (Hugging Face API), and the paper’s first version was submitted on 17 Nov 2026 (arXiv).Source: Hugging Face API and arXiv ·

The licence, read in full

  • The card’s licence tag is apple-amlr, and the LICENSE file in the repo is Apple’s Machine Learning Research Model licence. It grants a personal, revocable licence to use, copy, modify and distribute the model exclusively for Research Purposes, defined as non-commercial scientific research, and says that does not include commercial exploitation, product development or use in any commercial product or service. Derivatives must be labelled as such and carry an attribution notice. This is not legal advice.Source: LICENSE file in apple/FastFrames-L ·
  • Other licences named in the README: the repo’s other FastFrames code is under the Apple Sample Code License, and the LingBot-World v2 code and weights follow CC BY-NC-SA 4.0.Source: GitHub README ·

The paper’s claims, not tested by us

  • The paper is titled “FastFrames: Video Generation at Interactive Latency” (arXiv 2611.23226, computer vision, seven authors from Gene Chou to Vladlen Koltun). Its abstract says VAE decoding and KV cache updates account for over half of end-to-end latency, and the authors replace both with lightweight alternatives.Source: arXiv ·
  • Speed claims in the abstract: the 1.3B model needs 36 ms per frame on one B200 GPU, matching a state-of-the-art model that is nearly 8 times slower, and the 14B model needs 45 ms per frame on a node of 8 B200 GPUs, with better visual quality. The README’s own timing table for FastFrames-S on one GPU totals 35.4 ms.Source: arXiv ·
  • The project page repeats these figures and shows curated comparison clips. It gives no independent benchmark, and we did not use its per-clip numbers.Source: Project page ·

FastFrames-S, the smaller one

  • FastFrames-S has 1.3B parameters. Its repo holds only two small networks, a KV refiner and a pixel decoder; per the README it keeps the frozen two-step Causal Forcing++ generator and trains the networks around it, and the card’s base models are Wan2.1-T2V-1.3B and Causal-Forcing. The same repo also carries an optimized LingBot 1.3B that the README says is 16 times faster than the official model.Source: Hugging Face card (apple/FastFrames-S) ·

On this site

Not confirmed

  • A GPU or memory minimum for FastFrames-L, and whether it runs on a single GPU: the README states none, and its only L benchmark command uses 8 GPUs. The abstract’s 14B figure is also for 8 GPUs.Source: GitHub README ·
  • Image-to-video: not stated. VBench scores: the README points to a table that we did not read. Independent tests: we found none, and our search tool found no page about the model at all, only unrelated products with similar names.Source: GitHub README ·
  • Hosted versions: the card lists no inference provider. It does list three community Spaces (mrfakename/fastframes-l, akhaliq/FastFrames-L and linoyts/fastframes-l-live), which we did not open and cannot vouch for.Source: Hugging Face card (apple/FastFrames-L) ·

Sizes

FastFrames L file sizes, and this site

Our own sums from the Hugging Face file lists of apple/FastFrames-L and apple/FastFrames-S, read on 2026-11-19, in decimal GB. These are download sizes, not memory needs; no page states a memory figure. This site’s numbers come from its own settings.
-Size on Hugging FaceWhat it is
fastframes-l/dit.safetensors28.58 GBThe final 14B generator
fastframes-l/kv_refiner.safetensors1.78 GBThe KV refinement network
fastframes-l/pixel_decoder.safetensors0.29 GBThe pixel decoder that replaces VAE decoding
The three final files together30.6 GBWhat you need for generation (our sum, rounded)
fastframes-l-intermediate (3 folders)142.89 GBCausal teacher, consistency model and DMD model, for training and research
FastFrames-S: kv_refiner and pixel_decoder0.69 GB and 0.26 GBThe small networks only; the 1.3B generator is a separate model
This siteNothing to downloadOne photo edited in about 37 seconds, 10 credits (1024 × 1024); no video

Where to run it

How to run FastFrames L on your own GPU

  1. Check the licence first

    The licence file allows research use only and excludes commercial use and use in a commercial product or service. If your plan is commercial, FastFrames-L is not the model for it.

  2. Prepare a Linux machine with an NVIDIA GPU

    The README asks for Linux x86-64, an NVIDIA GPU and a CUDA 12.8-compatible driver, then uv sync. It states no memory minimum, and its L benchmark example uses 8 GPUs, so plan for more than one.

  3. Download the checkpoints and generate

    Follow the checkpoint guide in the repo, then run scripts/generate.py with --model ff-l, a prompt and a seed. The README’s example output is 832 × 480, 81 frames, 16 fps.

  4. Community Spaces: listed, not checked

    The card lists three community Spaces. We did not open them, and a Space run by someone else is not covered by this page’s facts.

  5. Need a still image instead? Use this site

    This site currently delivers images only, not video. Upload one photo and describe the change in the editor, up to 400 characters.

Prompts

Photo edit prompts for this site’s editor

These prompts are for this site’s photo editor, which edits still images with a different model from FastFrames. We did not run them for this page, and they do not make video.

One prompt per one-tap edit

  • Remove background clutter

    Edit my uploaded photo: remove distracting clutter from the background (stray objects, litter, bins, cables, signs and people in the background) and fill those areas naturally with what is around them. Keep the framing as it is. Keep every face, expression, pose and outfit exactly the same. No text, logos or watermarks.

  • Plain studio background

    Edit my uploaded photo: replace the background with a plain, softly lit light-grey studio backdrop with natural soft shadows. Keep the framing as it is. Keep every face, expression, pose and outfit exactly the same. No text, logos or watermarks.

  • Professional headshot

    Edit my uploaded photo into a clean professional headshot: a smooth neutral grey-blue background with a gentle gradient, soft even studio light on the face, the same crop or a slightly tighter head-and-shoulders crop. Keep every face, expression, pose and outfit exactly the same. No text, logos or watermarks.

  • Fix the lighting

    Edit my uploaded photo: fix the lighting only. Balance the exposure, lift dark shadows on faces, tame blown-out highlights and correct the white balance so skin tones look natural. Keep every face, expression, pose and outfit exactly the same. No text, logos or watermarks.

  • Warm golden-hour light

    Edit my uploaded photo: change only the light to warm, low golden-hour sunlight from one side, with soft long shadows and a warm glow, as if taken shortly before sunset. Keep the scene and framing. Keep every face, expression, pose and outfit exactly the same. No text, logos or watermarks.

  • Classic black and white

    Edit my uploaded photo into a classic black and white photograph with rich contrast, deep blacks, clean whites and fine film grain. Keep the composition exactly as it is. Keep every face, expression, pose and outfit exactly the same. No text, logos or watermarks.

  • Watercolor painting

    Repaint my uploaded photo as a soft watercolor painting on textured white paper, with loose washes and visible brush edges, keeping the same composition. Keep every person and pet recognizable, and do not add or remove anyone. No text, logos or watermarks.

  • Sharpen and clean up

    Edit my uploaded photo: gently sharpen it, reduce noise and compression artifacts, and recover fine detail in hair, fabric and edges, without a plastic or over-smoothed look. Keep every face, expression, pose and outfit exactly the same. No text, logos or watermarks.

Pricing on this site

Credit packs or a monthly plan.

Credits come in one-time packs ($4.99 for 50 credits, $9.99 for 120 credits, $19.99 for 300 credits); each image costs 10 credits. Or pick a plan from $9.90 a month (500 credits), cancel anytime.

  • Basic · $9.90/month500 credits a month · about 50 images · $0.20 per image
  • Pro · $19.90/month1500 credits a month · about 150 images · $0.13 per image

Billed every month and renews automatically. Cancel anytime in Settings; you keep the plan until the end of the month you paid for. Each month's credits are for that month; unused ones expire.

Refund policy

FAQ

Questions, answered

What is FastFrames L?

FastFrames-L is a 14B-parameter text-to-video model released by Apple, described in the paper “FastFrames: Video Generation at Interactive Latency”. It is an autoregressive, two-step video diffusion model that makes 480p video, distilled from Wan2.1-14B, and its weights are under a research-only licence.

Is FastFrames L a frame interpolation model?

No, as far as the model card, the README and the paper’s abstract say. They describe text-to-video generation, with a text prompt and a seed as input, and none of them mentions interpolating between frames of an existing video.

Is this site FastFrames, and can I make video here?

No. This site is not Apple, does not run FastFrames and currently delivers images only, not video. Its own editor uses a different model: upload one photo and write the change (up to 400 characters), or tap a ready-made edit.

What is the difference between FastFrames S and FastFrames L?

S has 1.3B parameters and L has 14B. S builds on Wan2.1-1.3B and a frozen Causal Forcing++ generator, and its repo holds only two small networks. L is distilled from Wan2.1-14B and ships a full 28.58 GB generator. The paper’s abstract claims L gives better visual quality.

Can I use FastFrames L commercially?

Not under the licence file in the repo, which grants use exclusively for research purposes and says that excludes commercial exploitation and use in a commercial product or service. Parts of the code use other licences. This is not legal advice.

How do I run FastFrames L, and how much VRAM does it need?

The GitHub README asks for Linux, an NVIDIA GPU and a CUDA 12.8-compatible driver, then uv sync, the checkpoints and scripts/generate.py with --model ff-l. It states no VRAM minimum, and its L benchmark example uses 8 GPUs. The final generator files total about 30.6 GB, summed by us from the file list.

How fast is FastFrames L?

The paper’s abstract claims 45 ms per frame for the 14B model on a node of 8 B200 GPUs, and 36 ms per frame for the 1.3B model on one B200. Those are the authors’ figures, not tested by us.

Can I try something here for free?

Partly: new accounts get 10 free credits after signing in, enough for 1 free image (once per account, valid for 7 days, while the day's new-account gifts last). Each image costs 10 credits, from $0.13 per image on a monthly plan.

Just need one photo edited?

Upload it, write the change, and download the result in about 37 seconds.

Edit a photo here

Independent site. Not affiliated with Apple, arXiv or Hugging Face, and this site does not run FastFrames or make video. Facts about the model come from the makers’ own pages, read on 2026-11-19; the authors’ claims are not tested by us. This is not legal advice.