← sources

Struggling with Moss

source: reddit (reddit:StableDiffusion)published 2026-10-19 07:28ingested 2026-10-20 02:38

retained text

Hey! so I heard Moss is currently THE best model now for uncensored realism. I tried it with n$fw prompts I copypasta'd from Kroma 0.3.1 and Krea 2 which were absolutely perfectly done on Krea/Kroma's side but the same prompts I put into Moss are outright shit.

While the adherence is perfect from a concept standpoint, the people in the images are always deformed, ugly, missing and melted limbs, fingers etc. the faces look older than they're supposed to as well as body type not registering via the prompt.

What am I doing wrong? I tried Comfyui default workflow, Moss SFT on:

\* 25 and 50 steps

\*euler/er\_sde simple/sgm uniform in respective combinations

\* with prompt enhancement and without, as prompt enhancement was pretty much only fixing the grammar on my prompt.

\* 2mp and 4mp

\* about 20-30 seeds on 2 n$fw prompts, absolutely no luck on either. one was just a scantily clad standing 1girl image ffs and it failed to do that massively.

I also tried sft on the official moss-image custom node workflow and getting just noise

Got a few things I wanna clarify:

\* SFT or base? especially for n$fw?

\* is the official comfy workflow bugged at all?

Please help me, I almost felt gaslit last night when I tried Moss. I just couldn't see literally any benefit on using Moss over Kroma, it felt worse than SDXL besides prompt adherence and I'm sure I'm doing something wrong.

document fetched 2026-10-20 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1x9qdcb/struggling_with_moss