Implemented EasyCache natively for (almost) all supported models in ComfyUI, and created LazyCache for a more generic and universally compatible implementation of the core concepts of EasyCache.

What is it?

EasyCache is in the family of step-skipping tricks like TeaCache that reuses cached values from a previous step whenever a step is determined to be unnecessary to sample. Unlike TeaCache, EasyCache only has 1 hyperparameter: reuse_threshold, and can quickly be adjusted based on model needs. It is also 'easy' to implement in that it only cares about the inputs and outputs of the models' forward functions, not anything inside them. I implemented it by making sure (almost) every supported ComfyUI model's forward function has DIFFUSION_MODEL wrapper compatibility - no internal model code was modified.

Each modified model type was tested with an example or template workflow with and without EasyCache enabled to make sure nothing that worked before broke.

Also invented a more universal but generally worse implementation of EasyCache I dub LazyCache - forget caring about caching individual conds surrounding the models' forward function, just deal with the full combined latents before and after each step. Sometimes when one works poorly for a model, the other does better - always favor EasyCache first though. To implement it, I added a PREDICT_NOISE wrapper handled by outer_predict_noise function in CFGGuider. There was no place previously to cleanly wrap around the noising of each sampling step before/after CFG, so this wrapper 'endpoint' can be potentially used for other things.

Usage:

EasyCache should ideally make the composition very similar to the non-EasyCache generation, except with a quality loss.

Here is what each parameter does on the nodes:

• reuse_threshold: value used to decide how liberally steps will be skipped. The higher the value, the more steps will be skipped (higher speedup, lower quality). The lower the value, the less steps will be skipped (lower speedup, higher quality). If you need to finetune the value and want to see what numbers the code sees, toggle verbose to True. Your console will be spammed by a bunch of log statements for each sampling step.
• start_percent: determines what relative point in the sampling steps to begin considering skipping steps. The early steps determine overall composition and are important in maintaining image quality. If you find that the resulting image strays too far from the non-Cached version for a particular model, consider increasing start_percent.
• end_percent: determines what relative point in the sampling steps to end considering skipping steps. The late steps are responsible for some of the finer details of the image. If you find the finer details lacking, consider decreasing end_percent.

Results:

Flux Kontext w/ 40 steps:

• EasyCache enabled - threshold: 0.2, start_percent: 0.15, end_percent: 0.95
  • Skipped 21/40 steps (2.11x speedup)
• LazyCache enabled - threshold: 0.2, start_percent: 0.15, end_percent: 0.95
  • Skipped 22/40 steps (2.22x speedup)

Qwen Image Edit w/ 20 steps:

• EasyCache enabled - threshold: 0.1, start_percent: 0.2, end_percent: 0.95
  • skipped 9/20 steps (1.82x speedup)
• LazyCache enabled - threshold: 0.2, start_percent: 0.2, end_percent: 0.95
  • skipped 9/20 steps (1.82x speedup)

WAN2.2 14B First Frame w/ 20 steps (10 per sampler):

• EasyCache First Sampler:
  • threshold: 0.10, start_percent: 0.15, end_percent: 1.0
    • skipped 3/10 steps (1.43x speedup)
• EasyCache Second Sampler:
  • threshold: 0.10, start_percent: 0.0, end_percent: 0.95
    • skipped 5/10 steps (2.00x speedup)
• LazyCache First Sampler:
  • threshold: 0.12, start_percent: 0.15, end_percent: 1.0|
    • skipped 2/10 steps (1.25x speedup)
• LazyCache Second Sampler:
  • threshold: 0.12, start_percent: 0.0, end_percent: 0.95
    • skipped 4/10 steps (1.67x speedup)