Auto Merger - wkpark/sd-webui-model-mixer GitHub Wiki

Overview

Please see auto-MBW and https://huggingface.co/Xynon/SD-Silicon

References

• Auto MBW v1
• Auto MBW v2 - AGPL
• https://github.com/THUDM/ImageReward
• https://github.com/christophschuhmann/improved-aesthetic-predictor
• https://github.com/grexzen/SD-Chad
• https://huggingface.co/cafeai/cafe_aesthetic

Image scoring classifiers

Select an image score classifier to score images

• aes-B32-v0
• cafe_aesthetic: https://huggingface.co/cafeai/cafe_aesthetic
• cafe_style
• cafe_waifu
• image_reward: ImageReward pip install ImageReward - https://github.com/THUDM/ImageReward
• laion-sac-logos-ava-v2: https://laion.ai/blog/laion-aesthetics/
• laion-avs1

and you can test classifiers by upload or drop a image and click "Test Image Score classifier" button.

Optimizers

It uses Hyperactive optimizer

Please see https://github.com/SimonBlanke/Hyperactive?tab=readme-ov-file#overview

Options

• Search Time Limit (min): Limit execution time in min. default: 10 min.
• Variable Search Blocks: Perform optimize on selected blocks. .e.g) IN01, IN02, M00, OUT04, OUT05,... (default: all blocks)
• Variable models: only selected models. e.g.) B. (default: all models)
• Search Upper, Lower, Max Limit: Limit variables range. e.g.) 0.4 given variable -> -0.2 + 0.4 ~ 0.4 + 0.2. minimum 0.0, maximum 1.0
• Warm Start, Early Stop are Hyperactive options

Interrupt, Reset

• click Interrupt button to stop the current optimizer job.
• click Reset Optimizer to remove the current optimizer's internal variables.

Notes

for old Model Mixer, AutoMeger only support MBW(Merge Block Weights) mode.

Optimizer Methods

Basics

Convex vs Non-Convex: https://nbviewer.org/github/SimonBlanke/hyperactive-tutorial/blob/main/notebooks/hyperactive_tutorial.ipynb#convex

Hill Climbing Optimizer

https://github.com/SimonBlanke/Hyperactive/blob/master/examples/optimization_techniques/hill_climbing.py

• Local Optimizer to maximize score.

Repulsing Hill Climbing optimizer

https://github.com/SimonBlanke/Hyperactive/blob/master/examples/optimization_techniques/repulsing_hill_climbing.py

• Local optimzer + "The repulsing hill climbing optimizer tries to improve how hill climbing solves non-convex objective functions."