Introduction
Welcome to my Sampling Method Tester Series, where I’ll be using the X/Y/Z plot script to test different sampling methods. The purpose of this test is to gain an understanding of how these sampling methods behave with different Checkpoints and Clip Skips, and to visualize how differently they read the prompt.
To ensure a fair test, I’ll be using the same prompt text and parameters for each sampling method, except for the Sampling method and Clip Skip. I’ll be testing all the different sampling methods provided in Stable Diffusion with 2 different Clip Skips (1 and 2), including:
Euler a, Euler, LMS, Heun, DPM2, DPM2 a, DPM++ 2S a, DPM++ 2M, DPM++ SDE, DPM fast, DPM adaptive, LMS Karras, DPM2 Karras, DPM2 a Karras, DPM++ 2S a Karras, DPM++ 2M Karras, DPM++ SDE Karras, DDIM, PLMS, and UniPC.
To make this X/Y/Z plot test as fair as possible, I’ll use the same prompt settings for all other Checkpoints. If I come up with new prompt ideas, I’ll try to use the same prompt for other Checkpoints as well. And if you have any interesting suggestions, you’re welcome to share them too.
In certain cases, when the pictures have too many strange artifacts, I’ll try to adjust the prompt settings just for that sampling method. All the Checkpoints, Textual Inversion, VAE, and LoRA I use for the X/Y/Z plot test will be provided on the page, so you can follow along with my testing process.
If you found this post useful or interesting, please don’t hesitate to share it with your friends and followers! It would mean a lot to me and motivate me to create more valuable content about this topic. Your support is greatly appreciated, and I’m always open to feedback and suggestions for future posts. Let’s keep the conversation going and explore the fascinating world of sampling methods together!
Links
Checkpoint (Model):
UnstableInkDream
Textual Inversion:
bad_prompt_version2
bad_prompt_v2
* They are slightly different, but you can use any what you like.
SD VAE:
blessed2.vae.pt
Testing
Prompt:
Disclaimer: This prompt I did find at “CivitAI“. I just don’t remember witch post or Checkpoint I was looking at. When I find it or if some of you have seen it, I will update that prompt credit!
Positive prompt:
hyper realism, (anime Makoto Shinkai:0.4), old shabby house in city street, building, sky, cloud, day, scenery, tree, blue sky, sign, no humans, power lines, house, english text, railing, wide shot
Negative prompt:
nsfw, sketch, duplicate, monochrome, (worst quality:2.0), (low quality:2.0), (blurry:2.0), horror, geometry, bad_prompt_v2, wooden, car
Parameters:
Steps: 28, Sampler: UniPC, CFG scale: 7, Seed: 607425775, Size: 768×768, Model hash: f036d7b40b, Model: unstableinkdream_v73, Clip skip: 2, SD VAE: blessed2.vae.pt
Result:
Correction:
As you can see, some of the sampling methods such as LMS, DPM fast, and PLMS have produced some artifacts in our initial test. However, we can still improve their performance by testing them with different settings.
We’ll keep the Clip Skip at 1 but increase the Sampling Steps for each sampling method. After some experimentation, we’ve found that LMS requires at least 80 Sampling Steps to remove artifacts and produce satisfactory results. On the other hand, DPM fast and PLMS look okay at 50 Sampling Steps.
By adjusting the Sampling Steps, we can improve the performance of these sampling methods and achieve better results. It’s essential to test and fine-tune the parameters of each sampling method to get the best possible outcome for our X/Y/Z plot test.
Stay tuned for more updates on our sampling method tester series! Cheers!
