chapter fourteen

14 Checkpoint Merging

 

This chapter covers

  • Understanding how checkpoint merging works to easily create new models
  • Using A1111 to quickly iterate on Stable Diffusion 1.5 checkpoint merges
  • Merging Stable Diffusion XL models using ComfyUI
  • How to create custom ComfyUI workflows that can save us a lot of time when iterating on models.
“This is a fun part, I’m gonna take titanium white, pull it out as flat as I can get it, add the least little touch of alizarin crimson to it. Least little touch. Just to warm it up a little, give it a little flavor.”

-- Bob Ross, The Joy of Painting

In this chapter we’re going to learn about checkpoint merging, a technique which allows us to quickly and easily combine two (or more) existing Stable Diffusion checkpoints to create an entirely new checkpoint. As we’ve progressed through learning about Stable Diffusion and other diffusion models like Flux, it seems that our VRAM and computing resource requirements have continued to increase with each chapter. In our last chapter on training Flux LoRA, we were able to get some fantastic results but could only reasonably train on a fairly high end GPU. Checkpoint merging provides an alternative option that allows us to create entirely new models very quickly with extremely low VRAM.

14.1 How Checkpoint Merging Works

14.1.1 Refresher on Neural Networks

14.1.2 Merging The Model Weights

14.2 Stable Diffusion 1.5 Merging with Automatic 1111

14.2.1 Merging Checkpoint in the A1111 Interface

14.2.2 Using Our Merged Checkpoint

14.3 Stable Diffusion XL Merging with ComfyUI

14.3.1 The ComfyUI Workflow

14.3.2 Generating our Images

14.4 Conclusion

14.5 Summary