Apple Releases FastFrames-S and FastFrames-L Text-to-Video Models

Apple Releases FastFrames-S and FastFrames-L Text-to-Video Models

Sample outputs are available on the model card.

At a Glance

Item Value
Repository apple/FastFrames-S
Publisher guide Apple: models and licenses
Published 2026-11-17
License apple-amlr
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Apple has released the text-to-video models “FastFrames-S" (1.3B) and “FastFrames-L" (14B) on Hugging Face. According to their model cards, both are autoregressive 2-step video diffusion models designed to generate 480p video with “interactive latency." These details were confirmed on the official Hugging Face repositories. Meanwhile, the paper is currently listed as “coming soon" on arXiv and has not yet been published.

Specifications

  • Parameter count: 1.3B for FastFrames-S and 14B for FastFrames-L
  • Base model: FastFrames-S is registered as a finetune of Wan-AI/Wan2.1-T2V-1.3B, and FastFrames-L of Wan-AI/Wan2.1-T2V-14B
  • Architecture: Autoregressive 2-step video diffusion model
  • Output: 832 × 480, 81 frames, 16 fps (listed as the output of scripts/generate.py)
  • License: Conditions regarding commercial use and other details are not specified in the current documentation. Please check the license text before use
  • Operating requirements: Linux x86-64, an NVIDIA GPU, and drivers supporting CUDA 12.8 are required
  • Training recipe: For FastFrames-S, the 2-step frame-level Causal Forcing++ generator is frozen, and the surrounding network is trained. For FastFrames-L, Wan2.1-14B is distilled into a 2-step causal model using the Causal Forcing++ recipe, while training a pixel decoder and KV refinement network. For L, it is reported that checkpoints have been released for the causal teacher trained on Wan2.1-14B latent representations, the consistency distillation model, the DMD distillation model, the final generator, the pixel decoder, and the KV refinement network

Recommended step counts, samplers, and CFG scales are not mentioned in the documentation.

Performance and Quality

The model cards provide the component-wise latencies for FastFrames-S (1 GPU, 480p, full KV cache). These figures were published by Apple.

Component Calls Latency per call Total latency
DiT evaluation, t=1000 1 16.1 ms 16.1 ms
DiT evaluation, t=833 1 16.2 ms 16.2 ms
KV refinement 1 2.2 ms 2.2 ms
Pixel decoder 1 0.9 ms 0.9 ms
Total 35.4 ms (paper: 36 ms)

Most of the 35.4 ms total comes from the two DiT evaluations, with the KV refinement and pixel decoder taking around 3 ms combined. However, the documentation does not clarify what unit of processing this latency represents. Speed comparisons with other models cannot be made from this table alone. Comparisons against baselines are said to be in the profiling guide referenced by the card, but that guide is not included in the current materials.

Additionally, the cards state that the evaluation procedure can reproduce the Quality, Semantic, and Total scores from “Table 1" of VBench. However, the exact score values are not provided in the documentation and are not presented here.

Furthermore, an optimized version of LingBot 1.3B has been released alongside these models. It is reported to be 16x faster than the official model and capable of smoother camera-conditioned manipulation. These figures are claims made by the publishers.

Strengths and Use Cases

Tagged as text-to-video, this model generates 832 × 480, 81-frame, 16 fps videos from text prompts. The model cards state that the objective is generation with “interactive latency," making it well-suited for workflows that require quick turnaround times for iterative experimentation or applications that return visuals based on user input.

The latency breakdown for FastFrames-S (total 35.4 ms) is as shown earlier. These figures are provided by the publisher.

The model cards also highlight camera-conditioned manipulation demos for LingBot as a use case. The optimized LingBot 1.3B is reported to be 16x faster than the official model, enabling smoother camera control. Demos that can be launched on either 1 GPU or 8 GPUs are provided, running on http://localhost:8000 after startup.

The changes from the original models are as follows:

  • The original models are Wan-AI/Wan2.1-T2V-1.3B (S) and Wan-AI/Wan2.1-T2V-14B (L)
  • Standard multi-step diffusion has been distilled into autoregressive 2-step generation
  • Additionally, a pixel decoder and a KV refinement network have been trained

According to the original Wan2.1 model cards, the 1.3B model targets 480p, while the 14B model supports both 480p and 720p. However, whether the same resolution options are available on the FastFrames side cannot be confirmed from the current documentation. The FastFrames cards specifically mention 480p output.

How It Differs from Similar Models

A previous article on our site, “LongLive-Plug-Wan2.1-T2V-14B-few-step": Generated Videos and Distribution Files List, covers another model built on the same Wan2.1-T2V-14B foundation. According to that article, that model is a LoRA adapter applied to the base model aimed at low-step generation, and requires the base model to run.

In contrast, FastFrames-L is released as a complete set of distilled causal model checkpoints rather than an adapter. It differs in that, alongside 2-step autoregressive generation, it also trains a pixel decoder and a KV refinement network.

Another video generation model, covered in “LongLive-Plug-Wan2.1-T2V-14B-cfg" Video Generation Model: Generated Videos and Distribution Files List, is an adapter that distills CFG into conditional-only generation, eliminating the need for an unconditional inference branch. That model alone does not provide a multi-step reduction effect. Because there are no materials directly comparing the speed or quality of both approaches, their relative merits cannot be determined.

Distributed Files

Weight files published in apple/FastFrames-S, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
fastframes-s/kv_refiner.safetensors 692MB
fastframes-s/pixel_decoder.safetensors 255MB
lingbot-1.3b/kv_refiner.safetensors 692MB
lingbot-1.3b/pixel_decoder.safetensors 128MB

How to Get It

The models are distributed in safetensors format. FastFrames-S is available on Hugging Face as apple/FastFrames-S, and FastFrames-L as apple/FastFrames-L. model_facts.gated is set to false.

To set up the repository, run the following command in the project root. The operating requirements are Linux x86-64, an NVIDIA GPU, and drivers supporting CUDA 12.8.

uv sync

The FastFrames core and third-party weights can be obtained following the checkpoint guide.

The command to generate a video from text is shown below. Use --model ff-l when using FastFrames-L.

uv run python scripts/generate.py --model ff-s --checkpoint-dir checkpoints \
    --prompt "A corgi surfing a wave at sunset" --seed 0 --output outputs/corgi.mp4

Latency can be measured with the following commands. An example with --gpus 8 is provided for L.

uv run python evaluation/latency/benchmark.py ff-s
uv run python evaluation/latency/benchmark.py ff-l --gpus 8

A script named scripts/generate_project_page.py is also provided to reproduce the 20+ videos featured on the project page.

The only supported tools listed are the scripts included in the repository and the LingBot demo. Compatibility with ComfyUI or Diffusers is not mentioned in the current documentation. Specific memory consumption figures during generation are also not provided.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — apple-amlr (Commercial use prohibited): Apple’s Machine Learning Research Model License. Limited to non-commercial scientific research; use in commercial products or services is expressly excluded.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

This model’s license (apple-amlr) does not permit commercial use. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Related Articles

What to Read Next

Sources