A temperamental LoRA trained on 51 hentai vids. Reading the instructions is highly recommended.

Weight

Doesn't have much of an effect at low weights. 1.0 is a starting point, 1.2 was used for most sample images.

Tags

The activation tag is hstyle, and h00 to h50 each corresponds to a unique source. I recommend including 1 or more of these, otherwise SD gets lost. Below are the combos used for the samples:

hstyle, hs03, hs08
hstyle, hs03, hs08, hs17, hs49
hstyle, hs00, hs04, hs11, hs19, hs37
hstyle, hs05, hs06, hs09

Other useful tags:

• anime coloring : Style reinforcement

• vum : Hashed token I found that produces a vintage style, some might enjoy

Negative tags:

• close-up : Training data consists of 16:9 images, sometimes taken from panning shots. This helps with output being extremely zoomed in. You can also add head out of frame as a negative prompt if you are desperate.

Dataset

51 videos were sourced from a "private collection" 😏, and I wrote a script to extract ~19 keyframes from each with a bias towards the middle of each video. This gave 969 (nice) frames which I pruned down to 621. Willing to release the script & outputs, not the full dataset ofc.