# MOSS-Image Some first impressions (this is bad) — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/d23a5a27efabdc27

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-19 00:24 ingested 2026-10-20 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1x9iqei/mossimage_some_first_impressions_this_is_bad/)

## retained text



```
I've tried several versions of the model, with 25 to 50 steps, after researching the best workflows and settings with AI, using official versions and unofficial ones, etc. What I get is slightly better prompt adherence, but poor images, with a lower IPR than SDXL and an even lower one than Z-Image. Honestly, I find it **almost as disappointing as Flux**. So I'll pass.

In terms of quality-to-speed ratio, here's my ranking from worst to best:

**Flux 1 < Moss Image < SD 1.5 < SDXL < Z-Image**

Why choose SD 1.5 over MOSS images? Because SD 1.5 is very fast, and out of 100 generated images, you can get one that is of truly high quality—and more faithful to the desired model—by using Dreambooth training (if I recall correctly), which is superior to LoRA. Afterward, you can simply enhance the image using a more recent AI.

The only exception is Flux 2 Pro, which is paid and only available through an API. If I could use that version, I'd rank it first, provided it were free and could run on regular hardware.
```



document fetched 2026-10-20 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1x9iqei/mossimage_some_first_impressions_this_is_bad
