Image Gen Caching Methods Documentation

[JohnLoveJoy]

1. What caching methods are

Caching methods in image generation frameworks (Stable Diffusion, SD XL, etc.) are techniques to store intermediate computations (activations, feature maps, etc) during the diffusion process so they can be reused in subsequent timesteps. This reduces total generation time, especially for high-resolution or multi-step images.

In short: caching accelerates image generation by avoiding repeated calculations.

2. Kobold recently Implemented the following caching methods:

- easycache

Description: Basic caching of reusable tensors. Supports threshold, warmup, start/end percentages.

Acceleration: Reuses intermediate activations to avoid recomputation.

Compatible models: Older SD pipelines and standard UNet-based models. Not optimized for DIT.

- ucache

Description: Unified cache manager that combines TEA-like, KV/attention, and intermediate tensor caching.

Acceleration: Reduces VRAM usage and keeps multiple caches in sync.

Compatible models: Both old SD XL models and DIT pipelines. Supports reuse threshold tuning.

- dbcache

Description: Stores activations in a persistent database (potentially disk-backed).

Acceleration: Avoids recomputation when GPU memory is limited.

Compatible models: Older SD XL models. Not specifically for DIT.

- taylorseer

Description: Advanced caching using residual/error heuristics to decide which activations can be reused.

Acceleration: Reduces unnecessary recomputation based on predicted error.

Compatible models: Older SD XL pipelines. DIT compatibility is experimental.

- cache-dit

Description: Caching designed specifically for DIT pipelines. Optimizes UNet activations and attention tensors during denoising.

Acceleration: Reduces recomputation in DIT multi-step diffusion.

Compatible models: Only DIT-based pipelines. Not for older SD XL models.

- spectrum

Description: Adaptive frequency-based caching. Classifies activations as hot/warm/cold. Supports warmup, λ, window size, flex window, and stop parameters.

Acceleration: Stores hot activations in GPU RAM and warm activations in system RAM. Reduces recomputation efficiently.

Compatible models: Newer SD XL pipelines and DIT variants.

3. Key cache parameters

reuse_threshold: minimum reuse score for caching activations

start / end: percent of timesteps to enable caching

warmup: number of initial steps before caching fully engages

Fn / Bn: compute block sizes for partial caching

spectrum-specific: w, m, lam, window, flex, stop (control adaptive hot/warm/cold caching)

4. Model compatibility summary

Old SD XL / UNet: easycache, ucache, dbcache, taylorseer

DIT pipelines: cache-dit, spectrum, ucache

5. How to use it in practice.

Depending on the model you are using, enter the model/parameters in the "default parameters" field in JSON format, as shown in the example below.

Caching methods work particularly well in SD XL, Anima, and Z-Image Base, as these models require many steps. So far, my preferred method in terms of speed and quality is Spectrum, which achieves nearly 2× faster performance in Anima.

[wbruna]

The docs are a bit outdated (spectrum now works for DiT too), but see:
https://github.com/leejet/stable-diffusion.cpp/blob/master-529-630ee03/docs/caching.md

The distinction is mainly between:

• UNET: SD1.x, SD2.x, SDXL (including Pony and Illustrious), and variants/distills (Inpaint models, Instruct-Pix2Pix, SSD1B, Tiny UNets, etc)
• DiT: SD3.x, Flux, Chroma, Wan, Qwen Image, Anima, Z-Image, Ovis Image

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

[JohnLoveJoy]

Just so you know, cache-dit doesn't work.

[wbruna]

Seems to be working here:

  "gendefaults": {
    "scheduler": "default",
    "sampler": "default",
    "cache-mode": "cache-dit"
  },

get_learned_condition completed, taking 0.05s
generating image: 1/1 - seed 1
CacheDIT enabled - mode: DBCache+TaylorSeer, Fn: 8, Bn: 0, threshold: 0.080, warmup: 8
  |==================================================| 16/16 - 1.32it/s

CacheDIT skipped 4/16 steps (1.33x estimated speedup)
sampling completed, taking 12.11s

[JohnLoveJoy]

Hmm. Did you activate it through the interface or via a command line?

[wbruna]

Interface; well, kind of. That's a chunk from the .kcpps file.