← sources

FIZGIG MOSS Image lora training success

source: reddit (reddit:StableDiffusion)published 2026-10-20 21:03ingested 2026-10-21 02:38

retained text

I was able to create a pretty accurate character lora successfully for 8 step lora workflow in comfyui. For me this is as close as I have been able to get to Krea2 lora.

Fizgig - 40 high quality 1:1 images with about 80% face and 20% body images

IMPORTANT - Use detailed captions describing parts of the photo you want to carry over, hair styles, poses, clothing, expressions and what not.

* Preset - MOSS-Image Fast
* Adaptive LR on
* Min LR - 5e-5
* Max LR - 1e-4
* Max epochs - 70
* Save every 5 epochs
* Seed - 42
* Context Lora - 1
* Target Megapixel 0.5 (if you can run and have patient I'm sure 1 MP works better)
* Network type - LoKR
* LoKR factor - 8
* Base SFT off
* Weight avg EMA - 0.98
* Detect problem images on
* Per-image adaptive LR on
* Warm up look outliers off
* Memory & precision all on auto and save state at end on

Fizgig preferences

* MOSS-Image Base DiT - moss\_image\_base\_bf16.safetensors
* FLUX.2 autoencoder - flux2-vae.safetensors
* Qwen3.5-9B text encoder - qwen3.5\_9b\_bf16.safetensors

Comfyui workflow - creates image in about 90 sec on my RTX 3080

* Unet - base int8 convrot
* Clip - 9b int8 convrot
* VAE - flux2 vae
* Skip layer guidance DiT - single layers=22
* Lora loader stack - 8 step lora with 1 strength & your character lora with strength 1
* KSampler - steps 12, CFG 1.5 (1-3 works), sampler res-multistep, scheduler simple, denoise 1
* Aspect 1:1 5.5 MP

document fetched 2026-10-21 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1xb44x1/fizgig_moss_image_lora_training_success