This was made mainly for slime girl . I have not tested other cases where transparency would make sense.

v5 update:

I went through and fixed most of the transparency & issues still present in the dataset. So now transparency should be better, make less patchy slime, and work for more things (like , , multiple penetrations). Note that there are still issues with high/from above angles, size differences for oral, and the orientation for some unique types. Using a big stack of loras and/or a lot of high tag weights will also make proper transparency significantly harder to get.

Getting proper transparency to work with slime girl is a pain. Most approaches either use x-ray (which looks weird for slime girls), do some complicated prompting workarounds, or just deal with low annoyingly low success rates.

The dataset is a mix of different resolutions, aspect ratios, slime colors, positions, and types. It should be pretty responsive to color and transparency in most cases. You may need other loras to get concepts such as , , multiple , etc. properly consistent though.

translucent should be the only thing needed to trigger it, but you might get better results with additional tags/phrases such as slime girl, (object) in slime deep , perspective, goo girl, monster girl, (color) skin (such as blue skin, green skin, multicolored skin, etc), (object) visible through body, etc. Some types might work better with specific angles (especially the types that are still WIP). If you are getting mismatched internal sizes you can try adding (disconnected :1.2) to your negative prompt.

I recommend pairing this with a dedicated slime girl LoRA (such as DaSiWa-Illustrious-Slime-Girl, Realistic Slime Girls, or Slime Girls - Concept) and a checkpoint that is able to handle concepts well such as DaSiWa Illustrious | Anime. While a strength of 1.0 should be safe, it's possible the image could end up overcooked with other LoRAs. Try setting it to ~0.7-0.8 if you are getting distorted faces and/or patchy slime. That should be enough to guarantee the concept in most cases, but a good LoRA combination with some good prompting can get it to work with as low as ~0.25.

My typical generation args are either:

steps: 37
cfg: 4.0
sampler: euler ancestral
scheduler: simple

or:

steps: 40
cfg: 4.0
sampler: uni_pc
scheduler: ddim_uniform

unless I have Config Skimming enabled then I set my cfg to 16 or 24 with the skimming cfg at 4 or 3 respectively.

If you are having issues with four or six fingers and you are still using v4 or earlier, try updating to v5 and reinforcing your negative prompt.