# Moss SFT INT8-convrot is just fine — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/dcb5262d7e4b39ff

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-07 01:30 ingested 2026-11-08 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xpu1e0/moss_sft_int8convrot_is_just_fine/)

## retained text



```
I'm not sure why people feel like Moss Base, BF16, or magic settings/workflows are necessary to get good results out of Moss. (My guess is that it's because the original default settings were suboptimal.) I personally feel like I get better, more prompt-adherent results out of SFT with a fairly simple workflow (see below), albeit with different settings from what ComfyUI had previously recommended.

If you have other approaches that work for you, great! I'm not trying to persuade anybody to give up what is working for them. But for people with more limited vRAM and/or who find themselves frustrated wrestling with the Moss Base model, a full 50 steps, or complex workflows, you don't necessarily need to do anything fancy.

My feeling is that you can achieve big improvements for photographic Moss images with:

* SA\_solver sampler and SGM\_uniform scheduler
* 25 to 50 steps (for some reason multiples of 25 seem better to me, but this could easily be confirmation bias)
* CFG 3.0 (go lower for more impressionist artistic styles)
* Skip layer guidance for single\_layer 22, start/end percent 0–0.5 or 0.25–0.75 (I find that skipping 24 leads to more plastic skin)
* Generation size ≥\~2MP, and
* Careful, detailed (but not purple) prompting

Note that all of these images were made using the same models, similar settings, and seed 0; I workshopped the prompts, but there was no seed cherry-picking.

Here are two version of my current workflow: a [simple one](https://pastebin.com/9hsidf3a) and [one with space for up to 3 LoRAs and an optional upscale step](https://pastebin.com/1DUp3H3y).

*Caveat: Krea 2 definitely understands some advanced compositions/concepts better than Moss, Moss is more prone to odd anatomy, and Moss's backgrounds can be more messy with more visual nonsense. But in terms of everyday photo realism I feel like Moss is better; and with respect to seed diversity, Moss (even SFT) blows Krea 2 out of the water. If you want less visual nonsense in the background, describe the background more thoroughly or do an inpainting step with low denoise on the background using Krea 2.*


```



document fetched 2026-11-08 00:27 · canonical https://reddit.com/r/StableDiffusion/comments/1xpu1e0/moss_sft_int8convrot_is_just_fine
