(all previews were done with seed=0, cfg=4.0, steps=35)

Trained in both T2I (100% generated with Moss) and I2I (single image input) with the goal of improving nsfw anatomy and pose understanding.

Please note that this lora is definitely not an overall improvement and in many cases loses against the base model, including identity preservation.

On I2I it was trained mostly on uncensoring tasks (30% of its global training data, censored-uncensored pairs obtained from gelbooru) with the prompt:

transform the image into an uncensored version of it. detect and clear out mosaic censoring, bar censors and gaussian blur. keep everything else the same.

but its not good at that so I may either train on top of this exclusively on that task or wait for something like Moss-Edit to come out (it helps a lot when the model already knows nsfw anatomy, ex: spread pussy). Sometimes you get better results when describing what to uncensor - this is specially true in very blurry/pixelated mosaic cases.

It was also trained on 'reference to sex' image pairs - mostly to prevent the model from losing its identity preservation capabilities - created with both Anima (using known characters + static style + reference lora) and Krea2+Identity lora.

The caption for those pairs were initially just very simple and short sentences and/or booru-like tags. Only later in the training, I've prepended this prefix:

create a new image with the character identity from <image1>.

From my tests it seems that prefix doesn't help at preserving identity (probably because it wasn't trained for long enough) but it does help in cases where the difference in pose/scene between the input and output is big.

I'm not satisfied with the results of this lora because ultimately is not great at anything in specific - thus I can't really recommend it but I'm sharing because why not.

Trained with musubi-tuner (dev branch), LR=2e-5, optimizer=adamw8bit, batch_size=1, gradient_accumulation_steps=2, batch sizes: 512, 768, 1024. All layers from blocks: 22~30 were removed after training which resulted in improved image quality in T2I mode without any major negative impact.