caching.md

April 27, 2026 · View on GitHub

Caching

Caching methods accelerate diffusion inference by reusing intermediate computations when changes between steps are small.

Mode	Target	Description
`ucache`	UNET models	Condition-level caching with error tracking
`easycache`	DiT models	Condition-level cache
`dbcache`	DiT models	Block-level L1 residual threshold
`taylorseer`	DiT models	Taylor series approximation
`cache-dit`	DiT models	Combined DBCache + TaylorSeer
`spectrum`	UNET and DiT models	Chebyshev + Taylor output forecasting

UCache caches the residual difference (output - input) and reuses it when input changes are below threshold.

sd-cli -m model.safetensors -p "a cat" --cache-mode ucache --cache-option "threshold=1.5"

Parameter	Description	Default
`threshold`	Error threshold for reuse decision	1.0
`start`	Start caching at this percent of steps	0.15
`end`	Stop caching at this percent of steps	0.95
`decay`	Error decay rate (0-1)	1.0
`relative`	Scale threshold by output norm (0/1)	1
`reset`	Reset error after computing (0/1)	1

The reset parameter controls error accumulation behavior:

reset=1 (default): Resets accumulated error after each computed step. More aggressive caching, works well with most samplers.
reset=0: Keeps error accumulated. More conservative, recommended for euler_a sampler.

Condition-level caching for DiT models. Caches and reuses outputs when input changes are below threshold.

--cache-mode easycache --cache-option "threshold=0.3"

Parameter	Description	Default
`threshold`	Input change threshold for reuse	0.2
`start`	Start caching at this percent of steps	0.15
`end`	Stop caching at this percent of steps	0.95

For DiT models like FLUX and QWEN, use block-level caching modes.

Caches blocks based on L1 residual difference threshold:

--cache-mode dbcache --cache-option "threshold=0.25,warmup=4"

Uses Taylor series approximation to predict block outputs:

--cache-mode taylorseer

Combines DBCache and TaylorSeer:

--cache-mode cache-dit

Parameter	Description	Default
`Fn`	Front blocks to always compute	8
`Bn`	Back blocks to always compute	0
`threshold`	L1 residual difference threshold	0.08
`warmup`	Steps before caching starts	8

Steps Computation Mask controls which steps can be cached:

--scm-mask "1,1,1,1,0,0,1,0,0,0,1,0,0,0,1,0,0,0,1,1"

Mask values: 1 = compute, 0 = can cache.

Policy	Description
`dynamic`	Check threshold before caching
`static`	Always cache on cacheable steps

--scm-policy dynamic

Spectrum uses Chebyshev polynomial fitting blended with Taylor extrapolation to predict denoised outputs, skipping entire forward passes. Based on the paper Spectrum: Adaptive Spectral Feature Forecasting for Efficient Diffusion Sampling.

sd-cli -m model.safetensors -p "a cat" --cache-mode spectrum

Parameter	Description	Default
`w`	Chebyshev vs Taylor blend weight (0=Taylor, 1=Chebyshev)	0.40
`m`	Chebyshev polynomial degree	3
`lam`	Ridge regression regularization	1.0
`window`	Initial window size (compute every N steps)	2
`flex`	Window growth per computed step after warmup	0.50
`warmup`	Steps to always compute before caching starts	4
`stop`	Stop caching at this fraction of total steps	0.9