Uploads from my experiments with 1) continued/resumed LoRA training combined with 2) RL training
moil is not intended to be a LoRA for a specific reason or purpose - it's intended to be a replacement for a full model finetune. So it aims to be well rounded.
I recommend using it on the base Z-Image Turbo model - not tunes.
Start by using moil only before mixing in other LoRAs.
If you're using Comfy you have to make sure it's up to date to be able to load PEFT format loras.
Description
FAQ
Comments (4)
This is interesting, seems really good for NSFW. It changes the image a lot compared to most ZIT loras.
Yes, the weight delta on this LoRA is quite large; and over time I'll attempt to improve stability to tame and refine it a bit as well. But it is indeed the point to continue to train the LoRA so that the large weight delta accomplishes the same outcome as continually training a full finetune.
This is a high quality lora. What weight and sampler/scheduler combo do you typically use with ZIT?
I find it burns quite easily at higher weights, especially if combined with other lora(s)
Yup, it's not all that stable - I'm still actively training it and pushing the weight deltas to where I want them to be; but that means it's not very refined yet.
I use fully vanilla ZIT diffusers pipeline for inference.