Training an NSFW LoRA based on SD3.5L has proven to be extremely challenging. This LoRA was trained intermittently over more than a month, reaching a total of 193,920 steps. Although it still falls short of ideal performance, it represents the limit of my efforts.

This LoRA, named ?????? Art, is designed to generate works in the style of realistic portrait photography and supports NSFW content.

Recommended settings:

model strength: 0.5–0.7, clip strength: 1.0,

CFG: 3.0–3.5, 25 steps.

Trigger keyword: NudityArt.

Due to the nearly intractable issue of edge artifacts inherent in SD3.5L LoRA training, the following guidelines are recommended for parameter settings:

1. Keep LoRA strength below 0.7.

2. Prefer square or vertical images (horizontal images, especially ultra-wide ones, will inevitably exhibit edge artifacts).

3. Reducing image dimensions to multiples of ?? can slightly mitigate edge artifacts in horizontal images (though not eliminate them completely).