# A few tips for MOSS — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/4114856fa9fe89aa

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-24 06:49 ingested 2026-10-24 10:59

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xe1nmu/a_few_tips_for_moss/)

## extract

There were quite a few different posts about MOSS last week, so I tried to compile some general tips for taking better photos:

\- The Base model is more flexible and versatile, while the SFT is more rigid.

\- There are several types of Turbo/DMD Lora, and there are different versions for the Base, SFT, and Kromoss models. It’s worth getting them all, because these Loras combine well with the models. You can find the 8-step Loras for each model on the Lodestone's [HF website](https://huggingface.co/lodestones/models).

\- In addition to the basic Euler/Simple combination, many people also use er\_sde and bong\_tangent. MOSS works with a wide variety of sampler/scheduler combinations, but the Turbo/DMD Lora files don’t work well with all of them.

\- Turbo/DMD Lora settings sometimes give better results at a strength of 0.7.

\- You can use cfg 1 in Turbo/DMD mode, but it also works with higher cfg values. With Turbo, I use around 1.5–1.7, but you can go up to 3; without Turbo, it’s worth trying values from 3 to 5.

\- The secret sauce: you can also improve the image by adjusting the values of the Uncond skip (SLG) node. This only works with cfg values greater than 1! Setting end\_percent to 0.5 or 1 also gives different results, and for the single\_layers values, try these: 22 or 22,24 or 22,24,26

\- If the model doesn’t understand something, it will generate odd results or weird body parts. In such cases, it’s worth rephrasing the prompt.

\- List of tags at the end of the prompt can also clarify or improve the result.

\- I’ve noticed that the higher the resolution, the better the image quality; 1.5 or 2 MP gives better results.

\- My personal tip: If you’ve installed the RGB Smart Seed Variance node, you can get even more exciting results from the model

\- And one more tip: you can use Image Sharpen (high\_pass/strength 0.5) from the KJ-nodes package to slightly sharpen the image.

## retained text



```
There were quite a few different posts about MOSS last week, so I tried to compile some general tips for taking better photos:

\- The Base model is more flexible and versatile, while the SFT is more rigid.

\- There are several types of Turbo/DMD Lora, and there are different versions for the Base, SFT, and Kromoss models. It’s worth getting them all, because these Loras combine well with the models. You can find the 8-step Loras for each model on the Lodestone's [HF website](https://huggingface.co/lodestones/models).

\- In addition to the basic Euler/Simple combination, many people also use er\_sde and bong\_tangent. MOSS works with a wide variety of sampler/scheduler combinations, but the Turbo/DMD Lora files don’t work well with all of them.

\- Turbo/DMD Lora settings sometimes give better results at a strength of 0.7.

\- You can use cfg 1 in Turbo/DMD mode, but it also works with higher cfg values. With Turbo, I use around 1.5–1.7, but you can go up to 3; without Turbo, it’s worth trying values from 3 to 5.

\- The secret sauce: you can also improve the image by adjusting the values of the Uncond skip (SLG) node. This only works with cfg values greater than 1! Setting end\_percent to 0.5 or 1 also gives different results, and for the single\_layers values, try these: 22 or 22,24 or 22,24,26

\- If the model doesn’t understand something, it will generate odd results or weird body parts. In such cases, it’s worth rephrasing the prompt.

\- List of tags at the end of the prompt can also clarify or improve the result.

\- I’ve noticed that the higher the resolution, the better the image quality; 1.5 or 2 MP gives better results.

\- best schedulers for many samplers:  
\-- Similarity in Version 1 when SLG end\_percent 0.5: simple, sgm\_uniform, normal, beta57, bong\_tangent  
\-- Similarity in Version 2 when SLG end\_percent 0.5: beta, linear\_quadratic,  sigmoid\_offset, power\_shift, radiance\_shift

\- My personal tip: If you’ve installed the RGB Smart Seed Variance node, you can get even more exciting results from the model

\- And one more tip: you can use Image Sharpen (high\_pass/strength 0.5) from the KJ-nodes package to slightly sharpen the image.

\- Here's a funny tip: ddim\_uniform produces some really funny results with any sampler.
```



document fetched 2026-10-25 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1xe1nmu/a_few_tips_for_moss
