You can refer to the video above for a demonstration of this workflow, including its effects and usage instructions.

Try Oline：https://www.runninghub.ai/post/2019376193790808065/?inviteCode=rh-v1401

5 mode Compare oline:https://www.runninghub.ai/post/2019746684867190785/?inviteCode=rh-v1401

When the MeanCache node is enabled for acceleration starting from step 3, it can achieve up to 100% speed increase, though this may lead to a decline in detail construction.

Generally, the Balanced mode is recommended to balance quality and speed. The turbo mode is suitable for quickly testing ideas, while custom is better suited for users already familiar with this node.

For an optimal experience, I recommend applying caching acceleration between 10% and 90% of the steps. This provides a good balance of speed while keeping quality within an acceptable range!

Next is a detailed analysis of the parameters.：

Core Principle of MeanCache: By caching and reusing results from similar denoising steps, it reduces the number of model computations, thereby significantly accelerating the generation process.

cache_device: The device used to store the speed cache.

- cpu: Saves VRAM. There is a slight data transfer overhead between CPU and GPU, which may slightly affect generation speed.

- cuda: Graphics Processing Unit. Offers faster access speed but consumes more VRAM.

rel_l1_thresh:

This parameter is the relative L1 threshold for skip decisions. Simply put, at each denoising step, the MeanCache node needs to decide whether to skip computation.

If the current latent variable is sufficiently similar to the cached latent variable, allowing the cached result to be reused, then it skips; otherwise, it proceeds with computation.

A lower threshold results in higher quality latent variables. A higher threshold leads to more noticeable speed improvements but may cause quality degradation.

The recommended value is around 0.2 to 0.4.

skip budget:

This is the maximum number of steps that can be skipped throughout the entire denoising process. For example, with a 50-step process and a skip budget of 10 steps, up to 10 steps can be skipped.

A larger value means faster speed (e.g., 0.3 indicates a 30% speed boost), but it will inevitably lead to a drop in quality.

start step:

The step at which the acceleration node begins to take effect.

end step:

The step at which the node stops caching and accelerating.

These two parameters control the range of steps during which the node is active, including the steps eligible for skipping. Therefore, it is advisable to apply caching acceleration primarily during the middle steps.

For instance, in a 50-step process, accelerating between steps 5 and 40 is ideal. The early steps should be preserved as much as possible to ensure the integrity of the image's basic structure.

The final few steps are for refining details. For high-quality portraits, sacrificing some speed in these later steps is acceptable.

If the generated images lack detail, you can configure the node to stop caching acceleration earlier.

enable pssp Simple Threshold Mode:

Used to filter or smooth certain noise or artifacts during image generation. For those prioritizing generation speed or the original effect, this can be turned off.

peak threshold:

This is the threshold for the filtering mentioned above. A higher value produces "cleaner" images but at the cost of lost details.

adaptive k:

An adaptive parameter k is a coefficient that controls the degree of algorithm adaptation, used to dynamically adjust the intensity of smoothing or processing.

debug:

Debug mode. Enable this when troubleshooting image anomalies or studying the node's principles. It outputs additional intermediate step information to the console.