FastFrames is Apple’s released implementation of a video-generation system designed to produce 480p video with interactive latency. It offers two autoregressive, two-step diffusion models: FastFrames-S, with 1.3 billion parameters, and FastFrames-L, with 14 billion. The model page also describes an optimized LingBot 1.3B system for camera-conditioned video control, reported to run 16 times faster than LingBot’s official version. For FastFrames-S, profiling on one GPU reports 35.4 milliseconds total for the listed generation components, close to the paper’s 36-millisecond figure. The supplied generation example produces 832-by-480 video, 81 frames long, at 16 frames per second.
The two versions follow different training paths. FastFrames-S keeps a frame-wise Causal Forcing++ generator frozen and trains supporting networks around it. FastFrames-L starts from Wan2.1-14B, which is distilled into a two-step causal model; its pixel decoder and key-value refinement network are then trained, with intermediate and final checkpoints provided. The release includes code and guides for installation, latency benchmarking, text-to-video generation, VBench evaluation, and an interactive LingBot demo. Running it requires Linux, an NVIDIA GPU, and a CUDA 12.8-compatible driver. FastFrames code uses Apple’s Sample Code License, while LingBot-World v2 code and weights retain their upstream CC BY-NC-SA 4.0 terms.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.