hellozicky/navicache

This repository contains the official code for NaviCache, a research method accepted at ICML 2026. Its purpose is to make AI video generation faster without sacrificing too much quality. Video generation models are slow because they process the same information through many computational steps on every frame, even when the underlying content has not changed much from one step to the next. NaviCache addresses this by tracking how much the internal representations are actually changing, and skipping computations when the change is small enough not to matter. The approach does not require any setup beforehand. It does not need a separate calibration dataset or pre-processing phase. Instead, it learns on the fly during inference, updating its own estimate of feature change as the video is generated. A lightweight estimator decides at each step whether to skip a processing block or compute it fully. This is controlled by a threshold setting that lets users trade generation speed for output quality. The code supports three video generation models: Wan2.1, HunyuanVideo, and Open-Sora 1.2. For each model, there are shell scripts that handle setup and launch, and the NaviCache entry point is a single Python script that gets copied into the official model repository. The README provides recommended threshold and alignment step values for fast, medium, and slow presets across all three supported models. Installation requires cloning both the NaviCache repository and whichever official model repository you want to accelerate, then running the provided launch script from within the model's directory. Environment variables control the key hyperparameters, so you can adjust them without editing files. This is a research release for AI practitioners and researchers working with video diffusion models. The paper and project page were still forthcoming at the time of release. The repository includes comparison videos showing NaviCache outputs alongside several competing caching methods on the same prompt.