hanjq17/Spectrum

Approximates denoiser latent features using Chebyshev polynomial basis functions fitted via ridge regression, enabling multi-step feature forecasting with non-compounding error bounds. Achieves 4-5× speedup on state-of-the-art diffusion models (FLUX.1, Wan2.1-14B) while maintaining sample quality, with training-free deployment across popular frameworks (Hugging Face diffusers, PyTorch). Supports image generation (Flux, SD 3.5, SDXL) and video generation (HunyuanVideo, Wan2.1) through Hydra-based configuration with tunable acceleration parameters.