How to train MOSS-Image LoRAs locally (8 GB VRAM): settings, tips and workflow
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.
Proof1 community thread
Key points
- 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).
- 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:
- 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).
Sources (1)
- [1]How to train MOSS-Image LoRAs locally (8 GB VRAM): settings, tips and workflowr/StableDiffusion (top, day) · Oct 20, 01:16 PM
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.
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).
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
- Oct 20, 2026MiniMax H3 - Really is the best open-source video model.
- Oct 19, 2026How good is FLUX.3 Klein at editing? A 12-task comparison that surprised me
- Oct 19, 2026Fixed the vertical colour stripes in MOSS-Image SFT with a 1.9 MB LoRA (extracted, not trained)
- Oct 18, 2026Comfy-Org/Fibo-2.0
- Oct 18, 2026Same prompt, 3 models: SDXL (WAI-Illustrious v17) vs Krea 2 turbo vs MOSS-Image-SFT (anime scenes, no LoRA)