Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.
Multimodal & Vision / Inference & Quantization

Apple Releases FastFrames-L and FastFrames-S for Interactive-Latency Video Generation

Apple has published FastFrames-S (1.3B) and FastFrames-L (14B) on Hugging Face, pairing two-step autoregressive video distillation from Wan2.1-14B and Causal Forcing++ with sub-40ms frame timings and mixed licensing terms.

Listen to this articleBrowser Voice
Ready
Official company logo or product image for apple-fastframes-l-interactive-video-generation
Official image or organization logo from huggingface.co; identification only, no endorsement.
Event / context: 2026-11-16. Source check: October 28, 2026.

Apple Releases FastFrames-S and FastFrames-L for Low-Latency Video Generation

Apple researchers have published apple/FastFrames-L on Hugging Face alongside an accompanying research preprint, FastFrames: Video Generation at Interactive Latency (arXiv:2611.23226), authored by Gene Chou, Guandao Yang, Fujun Luan, Lars Mescheder, Wei Dong, Stephan R. Richter, and Vladlen Koltun. The release packages two autoregressive, two-step video diffusion models—FastFrames-S at 1.3 billion parameters and FastFrames-L at 14 billion parameters—engineered to generate 480p video at interactive frame latencies.

Alongside the two core text-to-video models, the repository ships an optimized build of LingBot 1.3B that the authors report runs 16x faster than the upstream official model while supporting fluid camera-conditioned control. In standard text-to-video generation via the repository scripts, the models output 832 by 480 resolution clips spanning 81 frames at 16 frames per second. Distributed in Safetensors format, the release has already drawn early developer attention on Hugging Face, with three community Spaces referencing FastFrames-L even though no hosted Inference Provider currently serves the weights.

Two-Step Causal Architecture, Distillation Pipeline, and Latency Metrics

Rather than relying on dozens of iterative denoising steps, both FastFrames models operate as autoregressive, two-step causal video diffusion systems paired with specialized decoding and cache maintenance networks. For FastFrames-S, the authors keep the two-step frame-wise Causal Forcing++ generator frozen and train surrounding components around it, specifically a pixel decoder and a key-value (KV) refinement network.

FastFrames-L scales this strategy to a larger backbone derived from Wan-AI's Wan2.1-T2V-14B. According to the model card, Apple's team first distills Wan2.1-14B into a two-step causal model using the Causal Forcing++ recipe, then trains the pixel decoder and KV refinement network in the same manner as FastFrames-S. Notably for researchers studying distillation recipes, the release includes all four intermediate and final checkpoints: a causal teacher trained on latents generated by Wan2.1-14B, a consistency-distilled model, a DMD-distilled model, and the final generator paired with the pixel decoder and KV refinement network.

To substantiate its interactive-latency claims, the model card publishes component-level profiling for FastFrames-S running on a single GPU at 480p with a full KV cache. Under those test conditions, Apple reports two DiT evaluations—one at timestep t=1000 taking 16.1 ms and a second at t=833 taking 16.2 ms—followed by a single 2.2 ms KV refinement call and a 0.9 ms pixel decoder pass. Those stages sum to 35.4 ms per frame (rounded to 36 ms in the accompanying paper), illustrating how aggressive step reduction and lightweight decoding shift inference costs toward two predictable transformer passes.

Runtime Environment, Multi-GPU Setup, and Developer Workflow

Deploying FastFrames locally targets a specific workstation or server stack: Linux x86-64 equipped with an NVIDIA GPU and a CUDA 12.8-compatible driver. Dependency setup relies on uv sync from the repository root, followed by downloading FastFrames and third-party weights according to the project's checkpoint guide. Developers can run single-prompt generation through scripts/generate.py or batch-render showcase clips using scripts/generate_project_page.py by toggling between --model ff-s and --model ff-l.

For latency verification and interactive testing, Apple provides dedicated scripts. Profiling runs through evaluation/latency/benchmark.py, with the documentation illustrating single-GPU execution for ff-s and an eight-GPU invocation (--gpus 8) for ff-l. Similarly, the interactive LingBot demo at apps/lingbot/demo.py spins up a local server on port 8000 using either one GPU or eight GPUs via CUDA_VISIBLE_DEVICES, giving researchers and creative-tool builders a repeatable harness to test camera-conditioned generation locally before integrating the pipeline into custom tools.

Licensing Boundaries, Hardware Demands, and Verification Caveats

Despite its visibility on Hugging Face and transparent release of intermediate Wan2.1-14B distillation checkpoints, FastFrames is not a straightforward permissive open-source release. The Hugging Face model card carries the apple-amlr license tag, while the repository's License and Acknowledgments section specifies that FastFrames code is governed by the Apple Sample Code License. Meanwhile, LingBot-World v2 code and weights remain under their upstream CC BY-NC-SA 4.0 license, which restricts commercial use and mandates share-alike distribution, alongside additional third-party terms documented for Wan2.1, Self-Forcing, and Causal Forcing++.

Hardware and evaluation caveats also shape how teams can adopt the release. Requiring CUDA 12.8 on Linux x86-64 narrows immediate compatibility on older enterprise clusters, and the documented eight-GPU example commands for FastFrames-L highlight the compute footprint of running the 14B variant at interactive speeds. Finally, while Apple provides a VBench evaluation guide to reproduce Table 1's Quality, Semantic, and Total scores, those benchmark figures are not included directly in the model card excerpt, and both the 35.4 ms frame timing and the 16x LingBot speedup remain author-reported metrics pending independent third-party testing.

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-19.

Follow your curiosity.

Pass the signal.