Research
SpecKCast forecasts diffusion transformer outputs in frequency domain for faster generation
A training-free method that uses DCT coefficients and Kalman filters to skip transformer evaluations, reporting up to 6x speedup on FLUX.1-dev, Qwen-Image, Wan2.1-14B and HunyuanVideo with under 0.05 GiB extra memory.
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
SpecKCast: Spectral Kalman Forecasting for Diffusion Acceleration