DeepCache for SD Forge

This is an implementation of DeepCache: Accelerating Diffusion Models for Free for SD Forge. The core is basically laksjdjf's ComfyUI node, with a few of my own improvements.

Only tested with SDXL, but should work with other models.

Installation

git clone https://github.com/LoganBooker/sd-deepcache.git into stable-diffusion-webui-forge/extensions/ folder.

Warning

The extension patches the Unet forward pass, so other extensions that also do this might not play nicely. There is no easy fix as far as I'm aware.

Improvements

• Keyed-cache dictionary to handle situations where the batch size changes during diffusion. This fixes errors such as Expected size 2 but got size 1 for tensor number 1 in the list, which is common when other extensions manipulate conditional/unconditional batching, such as Perturbed-Attention Guidance.
• Cache interval power curve to increase the interval over time. A hard value to tune, but 4 works well when applying the cache from start to finish.
• Forcing the cache on after a certain sigma, which helps to avoid one or more costly and usually unneeded caching steps near the end of diffusion.

Parameters

• Enable DeepCache for passes: Passes to apply DeepCache. None, first pass only, hires pass only, or both (recommended).
• Cache interval power curve: Optionally increase the cache interval over time. Starts at 1 and increases to Cache interval following a power curve. Set to 0 to disable (default).
• Force cache after sigma: Sigma to ignore the cache interval and always use the cache. Avoids costly caching steps when noise levels are very low. Default is 0.1.
• Cache start sigma: Sigma to enable the cache. Higher values enable caching sooner, but will reduce quality. Values between 5-1 are recommended, if using default sigmas ([0.29, 14.6]).
• Cache end sigma: Sigma to disable the cache. This should be lower than Cache start sigma. I'd recommend leaving this at 0, as disabling the cache deep into the diffusion process offers few benefits (at least for SDXL).
• Cache interval: Frequency of cache steps. Lower values (2-3) are slower but higher quality. Values above 5 are usually too destructive, but might not be bad for high step counts (50+).
• Cache depth: Layer to apply caching. Lower values are faster, but reduce quality. For SDXL, a value of 4 or higher is recommended. Default is 5.

Hires parameters

For the hires pass, all the above settings are not used, except for the cache interval and cache depth.

• Use first pass settings: Ignore the values specified in the hires section and use the Cache interval and Cache depth from the main parameters.
• HR Cache interval: if Use first pass settings is disabled, use this value for Cache interval.
• HR Cache depth: if Use first pass settings is disabled, use this value for Cache depth.