Unpolished Latex Catsuit

This LoRA is designed to create a more natural-looking latex catsuit, featuring a subdued, matte-like shine and visible overlapping glued seams for added realism.

I’d love to see the creative images people generate with this LoRA, let your imagination run wild and share your creations!

Tips:

• Strength 0.9–1.0 gives the best results for a "dull shine" and fine details.

• Steps 40 (you can use less but the latex won't look so nice).

• To emphasize the "dull shine," consider using descriptive prompts generated by a language model. For example:

"The rubber material has a dull, matte shine, reflecting soft light but without an overly glossy finish. The suit clings closely to the body, showing subtle reflections along the curves and edges, but the light is diffused, creating a smooth, understated sheen rather than sharp reflections. The lighting in the room is gentle, enhancing the rubber’s texture by highlighting its slight gloss while keeping a balanced, muted look. The overall effect of the outfit is sleek yet controlled, maintaining a refined, non-flashy appearance."

Known Issues:

• This LoRA doesn’t mix that well with others. While it works with person-specific LoRAs, the latex texture/shine quality might not be optimal (and I’m admittedly quite particular about that).

• The training dataset didn’t include front-zip catsuits, only rear-zip or neck-entry designs. If you specifically prompt for a front zip, you might get unusual results.

• A few dataset images featured latex toe socks (with individual toes visible), which can sometimes appear in your outputs even if you prompt for shoes, or worse, hybrids of toe socks and shoes. If I revisit this LoRA, I'll crop the toe socks out of the dataset images before training.

• It sometimes gets a bit over-creative with the glued seams on the suit, so they appear in strange places.

Training Details:

This LoRA was trained using the Civitai LoRA Trainer with the following parameters:

• Images: 32 (significantly reduced from earlier attempts for better focus on the theme)

• Epochs: 20 (with the best results from Epochs 16–20; final model based on Epoch 17 after extensive testing)

• Repeats: 13

• Resolution: 1024

• Optimizer: Prodigy

• Steps (for the Epoch I selected): 1768

• Images were captioned using Joycaption, removing any references to latex catsuits.

This was my sixth attempt at training this LoRA. The key to this version’s success was refining the dataset to include only images closely aligned with the desired theme. Interestingly, earlier attempts used three times as many images but delivered poorer results, proving that quality trumps quantity for Flux Dev LoRAs.

Disclaimer: I’m no expert, just someone who enjoys generating images. I rarely upscale or inpaint, so take my insights above with a pinch of salt.