# A potential fix for MOSS training issues.... — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/d23995e6deb63d73

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-19 01:53 ingested 2026-10-20 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1x9khtm/a_potential_fix_for_moss_training_issues/)

## retained text



```
**Update Monday Morning:** TLDR: The 60 setting worked very well.  Setting it at 30 which I've run since works even better, a good balance where it still learns composition but attends to finer detail at a similar level as Krea 2 and Qwen Image 2.1 training schedules.  Detail of all of this can be found below.

**Monday Afternoon EDIT:** 30 is the best setting so far for me. I've done trains today at 60 and 30, and 30 pulls ahead after about 300-400 steps and stays there, with some of the best likeness I've had from this dataset on any model and it translates to comfyui also. It also keeps a better balance of learning composition. It's also exciting to think about using this for composition-only trains at very low values, below MOSS's native 11%. I may open this control up to the other models in Fizgig. Fizgig's default is now 30.


**Kromoss** After a lot of testing, as good as Kromoss is in some ways , a couple of interesting points about using it as a base for training.  1) Loras seem to collapse with it at base within 700-800 steps, which rules it out for training.  2) It does not work well as a preview model either, as it renders the loras too differntly.  3) I have some ideas of making a 'Context lora' to use when training a lora you want to use in Comfy with Kromoss - I'll come back on this soon.


------Original post below this line------

So it's 2:43 in the morning here as I write this. I went to bed, then got up again, because something from my MiniMax H3 training wouldn't leave me alone - on H3 I had to shift the share of training steps at the clean end of the noise range (where faces and fine detail get refined) before likeness came good.

So I compared MOSS's training sigmas with Krea 2's and it was like looking at MiniMax all over again. MOSS's own schedule puts only about 11% of steps at the clean end (sigma below 0.5), Krea 2 is 18-28%. The rest go to layout and composition, which explains the big jumps epoch to epoch and why it's been so hard to nail the likeness.

I've just pushed an update (v8.7.0) with a new **Clean-end share** option for MOSS training, on by default at 60%. Clear the box to go back to the model's own ~11%.

My test run is going as I write - previews are holding their composition between epochs, and at epoch 12 it had already passed the best likeness my previous settings got to in 30 epochs.

**Numbers so far** - same 53-image dataset, rank 16, 0.25 MP, LR 1e-4, same seed, ArcFace likeness vs 3 reference photos:

| Epoch | 1 | 3 | 6 | 9 | 12 | 15 | 30 |
|:--|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
| Old schedule (~11%) | 29% | 40% | 44% | 52% | 58% | 63% | 67% (its best) |
| Clean-end 60% | 20% | 50% | 63% | 66% | 68% | 69% | still running |

So 60% passed the old run's best at epoch 12 (about 640 steps vs 1,590), and it's been sitting at 68-69% since rather than bouncing 10 points epoch to epoch.

60 is an experimental figure (and only set that high as I wanted definitive proof either way), so please play with it. As a community we can find the right setting - 60 as a setting will probably come down. 30 (about where Krea 2 sits) is a sensible next try. Post what you find here.  Whatever the case only 11% of steps attending to detail where likeness lives in training is too low.

[Fizgig on GitHub](https://github.com/shootthesound/Fizgig) · [v8.7.0 release notes](https://github.com/shootthesound/Fizgig/releases/tag/v8.7.0)


**Update - the 60% run has finished (70 epochs)**

Summary excellent and steady growing likness to the eye thoughout the run backed up by the numbers below:

53 images, rank 16, 0.25 MP, LR 1e-4, same seed, ArcFace likeness vs 3 reference photos.

| Epoch | 3 | 6 | 9 | 12 | 15 | 20 | 30 | 40 | 50 | 60 | 70 |
|:--|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
| Old schedule (~11%) | 40% | 44% | 52% | 58% | 63% | 60% | 67% | - | - | - | - |
| Clean-end 60% | 50% | 63% | 66% | 68% | 69% | 67% | 71% | 71% | 69% | 74% | 76% |

The old run stopped at 30 epochs, where it peaked at 67%. The 60% run was ahead of it at every epoch from 3 on, was still climbing after epoch 30, and finished at 76% (also hit at epoch 66).

It's steadier too. From epoch 10 to 29 the old run swung between 51% and 66%, about 2.4 points per epoch on average. The 60% run stayed between 63% and 70%, about 1.5 points per epoch.

```



document fetched 2026-10-20 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1x9khtm/a_potential_fix_for_moss_training_issues
