Version 1.1 Update.

This is a significant update with an expanded dataset and re captioning. I am still testing the epoch outputs but this I currently feel is the best balance without overcooking the concept.

It play reasonably well with other Loras although as usual, adjusting the weightings to best suit is needed.

I created the majority of my images using ComfyUI using the following paramaters

Max_Shift: 0.70

Base_shift: 0.20

Flux Guidance: 1.9-2.2

Sampler: Euler

Scheduler: Beta

Steps: 30

I have had trouble getting CivitAI's generator close to the accuracy and detail of my own images. Not sure on the sampler they use.

A more focused Lora focusing on Custard Sploshing.

The Lora has been training on 300 WAM/Sploshing images that I've curated from my collection. The image set is exclusively Custard scenes with the subjects wearing a variety of outfits, costumes in various poses and locations to give a varied data set.

The Lora works best with natural language prompts. There are are no trigger words needed for the lora. Simply specify the subject location, pose and substance is enough. Try to describe how the substance specifically interacts with the subject for the best results.

As far as settings go, The lora works well for strengths 0.75 to 1

I find it helps to give random names, especially if you want multiple people.

I've been sampling with guidance of 1.9-2.2 to test, simple euler schedule with 30 steps.

The usual samplers should work well Euler Simple/Beta and Deis/ddim_uniform are my go to samplers. I usually run 30 steps, but 20-40 is fine too.

It is trained with the full precision t5/clip conditioning. I have tested FP16 and FP8; both give excellent results obviously depending on the seed used.

Feedback welcome. and any buzz donations are appreciated.