sd-webui-bayesian-merger

What is this?

An opinionated take on stable-diffusion models-merging automatic-optimisation.

The main idea is to treat models-merging procedure as a black-box model with 26 parameters: one for each block plus base_alpha (note that for the moment clip_skip is set to 0).

We can then try to apply black-box optimisation techniques, in particular we focus on Bayesian optimisation with a Gaussian Process emulator.

Read more here, here and here.

The optimisation process is split in two phases:

1. exploration: here we sample (at random for now, with some heuristic in the future) the 26-parameter hyperspace, our block-weights. The number of samples is set by the

--init_points argument. We use each set of weights to merge the two models we use the merged model to generate batch_size * number of payloads images which are then scored.

2. exploitation: based on the exploratory phase, the optimiser makes an idea of where (i.e. which set of weights) the optimal merge is.

This information is used to sample more set of weights --n_iters number of times. This time we don't sample all of them in one go. Instead, we sample once, merge the models,

generate and score the images and update the optimiser knowledge about the merging space. This way the optimiser can adapt the strategy step-by-step.

At the end of the exploitation phase, the set of weights scoring the highest score are deemed to be the optimal ones.

Juicy features

- wildcards support

- TPE or Bayesian Optimisers. cf. Bergstra et al. 2011 for a comparison

- UNET visualiser

- convergence plot

OK, How Do I Use It In Practice?

Head to the wiki for all the instructions to get you started.