# How to train MOSS-Image LoRAs locally (8 GB VRAM): settings, tips and workflow — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/8ca4285a59fe867a

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-20 13:16 ingested 2026-10-21 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xarw4c/how_to_train_mossimage_loras_locally_8_gb_vram/)

## retained text



```
A lot of people have been asking me for the "right" settings to train MOSS-Image LoRAs, and some are getting terrible results after training for thousands of steps. So here's everything in one place, with screenshots of a real run.

**TL;DR:** for a character LoRA, start with the default settings and train **800 steps** (just an example: other types of training may need other settings). The likeness should already be there by step \~800; go to **1200–1500** only if your dataset is large or the concept is hard. If you're at 8000 steps, you've gone way past the point where the LoRA was good.

# The tool

I train with **AcademiaSD LoRAlab Trainer Studio**, a free, open-source local trainer (one installer, one launcher, NF4 models) that also does Qwen-Image, FLUX.2 Klein, Z-Image, Ideogram 4, SDXL, LTX-2.3 and more:

* GitHub: [https://github.com/AcademiaSD/AcademiaSD\_LoRAlab-TrainerStudio](https://github.com/AcademiaSD/AcademiaSD_LoRAlab-TrainerStudio)
* Video with the ComfyUI workflow: *"AMAZING MOSS-IMAGE! UNCENSORED, JSON + BBOX EDITING, LONG TEXT AND MORE"*: [https://www.youtube.com/watch?v=nXqRnyRUwgU](https://www.youtube.com/watch?v=nXqRnyRUwgU)
* No GPU? There's a ready-made RunPod template (link in the repo README).

MOSS-Image (OpenMOSS, Apache 2.0) is a 5B single-stream model. In Trainer Studio the transformer is NF4, so **training at 512² with rank 8 uses under 6 GB of VRAM** (\~1 s/step on an RTX 5080). 8 GB cards are perfectly fine.

# Settings that work (character LoRA)

https://preview.redd.it/3f2swpqchmwh1.png?width=2530&format=png&auto=webp&s=9b6c93fb767331814cba13ebad8c6a6ddd884a14

|Setting|Value|
|:-|:-|
|Base model|**Base** (the one OpenMOSS publishes for LoRA training; SFT also works)|
|Steps|**800** (1200–1500 for big datasets or harder concepts)|
|Learning rate|0.0004|
|LR scheduler|cosine, 100 warmup steps|
|Rank / alpha|8 / 8|
|Batch size / grad accumulation|1 / 4|
|Training resolution|512²|
|Save every|100 steps|

>⚠️ **These settings are just an example**, a starting point that works for this character LoRA. Each type of training (characters, objects, styles, concepts) and each dataset may need different settings: steps, learning rate, rank, resolution or captions. Run your own tests and compare checkpoints.

**Why 8000 steps goes wrong:** with gradient accumulation 4, every step already sees 4 images. 800 steps = 2,000 image views; 8000 steps = 32,000. At that point the LoRA has long since overfit, and adding more images doesn't fix that. More steps is not more quality here.

# How to pick the right checkpoint

1. Train with the defaults; a checkpoint is saved every 100 steps.
2. Check the previews, but **judge the final result in ComfyUI**: ComfyUI gives better quality and closer likeness than the in-trainer previews (they're there to follow the training, not to judge it).
3. Test 300 / 400 / 500 in ComfyUI. If the likeness isn't there yet, continue to 1200–1500 and test those checkpoints.

https://preview.redd.it/1y5va8w2hmwh1.png?width=2540&format=png&auto=webp&s=7f85fd1cbc97df033dd621c6eeaec815699ce8c9

If your results still look bad at 800–1500 steps, post your settings and a few dataset images (or open an issue on GitHub) and I'll take a look.

# Captions and dataset

* Use the built-in captioner (Qwen3-VL-8B NF4). It writes detailed prose following OpenMOSS's own recaptioning guide, with the trigger word first.
* Captions matter more than dataset size: a clean, varied, well-captioned dataset beats a huge one.

# ComfyUI

The LoRA is exported with the original layer names, so a recent **ComfyUI (nightly, with native MOSS-Image support)** loads it with the standard LoRA loader. The workflow I use is in the video linked above.

https://preview.redd.it/b6ee8ui7hmwh1.png?width=2554&format=png&auto=webp&s=487467e165f6755417bdf88e822b191b9d8b5f68

# 

If your results still look bad at 800–1500 steps, post your settings and a few dataset images (or open an issue on GitHub) and I'll take a look.
```



document fetched 2026-10-21 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xarw4c/how_to_train_mossimage_loras_locally_8_gb_vram
