Spectrum speeds up image and video diffusion 3.5x without quality loss

Spectrum is a ByteDance framework that accelerates diffusion-based image and video generation by predicting future steps over computing them all. It can be stacked on top of existing workflows (Hunyuan Video or Flux) to deliver similar outputs with fewer sampling steps.

In tests, Hunyuan Video clips generated with 50 steps are visually indistinguishable from Spectrum-accelerated versions using just 14 steps. This made for a roughly 3.5x speedup.

For Flux image generation, Spectrum maintains comparable quality while being 4.7x faster than the baseline. It’s also more accurate than competing accelerators like TaylorSIRE, which introduce artifacts and color shifts.

Spectrum relies on Chebyshev polynomials to model how features evolve over time, forecasting later steps cheaply. The team has already released code and instructions on GitHub, with support for multiple image (Flux, SDXL) and video models (Hunyuan, Wan 2.1).