BRKN BDSM V2 (LoRa) – What it is and what it’s trained to do

BRKN BDSM V2 is a restraint-and-position focused LoRa built to capture the emotion, tension, and geometry of BDSM practices without centering the sexual act itself. It was trained on 135 images, for roughly 7,000 steps at batch 32, with the goal of teaching the model the logic of restraint systems and the body language that comes with them.

The training emphasis is on restraints and materials: ropes and shibari, ladder/strap restraint setups, collars, wrist/ankle/waist/thigh cuffs, metal cuffs, chains, pipes, rusted metal elements, and accessories such as gag balls. Color is kept grounded and natural, with a palette that leans into blues, greens, blacks, and realistic lighting. Settings vary between indoor and outdoor, including wooded locations, industrial spaces, caves, and dungeon-like environments, so the same LoRa can pivot mood and atmosphere depending on your prompt.

For poses and configurations, BRKN BDSM V2 is designed to handle complex constraint logic: suspended and inverted positions, suspension by ankles, ankles-to-elbows ties, elbows-to-knees, collar-to-ankle connections, all-fours, folded, straight, spread-eagle, and other contortion-forward setups. Because these are mechanically complicated, the best results come from extra-detailed prompting (especially for limb placement and connection points). When prompted carefully, the LoRa helps keep anatomy readable and body parts present.

The dataset comes largely from the my own work as a photographer, with the appropriate rights to train and publish the LoRa. It’s meant to be flexible: black-and-white or color both work well, and it pairs nicely with Z-Image as a base model for realism.

Recommended settings: strength 0.6–0.8 for most use cases, with 1.0 as a hard ceiling (above that, you’ll start to see mutation depending on sampler/scheduler). For maximum realism, and personnel preference it was tested with Res2 + BumpTangent at 8,10, 1 CFG, and 1 noise.