My attemp to interpolate models with unet 686 weights and text_encoder 197 weights seperatedly. The goal is to extract MeinaMix's style, extract Counterfeit's background and PastelMix's intricate detail (but exclude their messy).

Manually decide each of these weights are impossible, I've introduced "RMHF" or Reinforcement Model-merging from Human Feedback.

https://github.com/TkskKurumi/DiffusersFastAPI

I'm naming it like a machine learning method XD, but in fact the method very simple. (I know barely nothing about Reinforcement Learning)

Let's say the merging composition is W, meaning model-index i, weight-index [j] has W[i][j] contribution. In each iteration, generate a random (not purely random, scale is scheduled) vector ε has the same shape as W, show user the pair of W-ε and W+ε generated images, select either and update W = W±ε.