After a very long while I got back into AI image stuff, so here is my MOSS Image LoRA for Fizz :)

Recommendations:

• CFG: 5.0

• Steps: 50

• Sampler: er_sde

• Scheduler: simple

• Strength: 1.0

• Resolution: Anything between 1024 and 2048. Lower and higher often creates problems with distortions and merged limbs.

For the prompt it is important to mention certain things to come out right.

Trident:

Mention "a trident with three curved prongs". If you only mention a trident, it will usually only generate two prongs. Don't mention the trident and it will be left out.

NSFW:

The LoRA is trained with the key word "cetacean_penis". However it often still generates a penis that is more resembling a traditional penis and it also tends to create something like a scrotum. Use the negative prompt and add "scrotum" or "glans". It also helps to use things like "tapering" in the positive prompt.

Gills:

Sometimes the gills seem to look like simple red stripes. You can mention "gill slits" and it will look a lot more like actual gills.

The LoRA seems to be rather versatile with styles but depending on the prompt details, it tends to easily go into the digital painting and semi-realistic direction. Be aware of that and prepared to play around a bit to find the right prompts.

Example images are all oneshot prompts (no in-painting, no regional prompt, no masks, etc.). Most of them only went through a hi-res second pass at 1.5x upscale.

I will try to fix a couple of issues I noticed with a second version later :)