Nodes/Eric_Qwen_Edit_Experiments/Eric Qwen-Edit Spectrum Accelerator
ComfyUI Node

Eric Qwen-Edit Spectrum Accelerator

Spectrum Accelerator — training-free ~3–5× diffusion sampling speedup. Predicts transformer outputs on skipped steps using Chebyshev polynomial regression instead of running all 60 transformer blocks. Wire between Loader and any execution node.

By EricRollei·Created 5 months ago·Updated 3 months ago· 19
Eric Qwen-Edit Spectrum Accelerator
  • pipeline
  • pipeline
enabletrue
warmup_steps3
window_size2
flex_window0.75
w0.50
lam0.10
M4
min_steps15
CategoryEric/QwenEdit

Inputs (9)

NameTypeDefaultDescription
pipelineQWEN_EDIT_PIPELINE
enableBOOLEANtrueEnable/disable Spectrum acceleration.
warmup_stepsoptINT31–20Number of initial denoising steps that always run the full transformer. More warmup = better forecaster initialization but less speedup. 2–4 recommended.
window_sizeoptINT22–10Base period between actual transformer evaluations (in steps). window_size=2 means every other step is cached after warmup. Higher = more aggressive caching.
flex_windowoptFLOAT0.750–2Window growth rate — the window_size increases by this amount after each actual eval. Later diffusion steps change features less, so larger windows are safe. 0 = fixed window.
woptFLOAT0.500–1Blend weight between Chebyshev predictor (w=1) and Newton forward-difference predictor (w=0). 0.5 = equal blend (recommended).
lamoptFLOAT0.100.001–10Ridge regularization for Chebyshev regression. Higher = smoother predictions, lower = more responsive to recent data.
MoptINT41–8Chebyshev polynomial degree. Higher M captures more complex feature trajectories but risks overfitting with few data points.
min_stepsoptINT155–100Spectrum auto-disables when num_inference_steps < min_steps. Low step counts (e.g. 8-step lightning LoRA) don't benefit.

Outputs (1)

NameTypeDescription
pipelineQWEN_EDIT_PIPELINE