A temperamental LoRA trained on 51 hentai vids. Reading the instructions is highly recommended.<h2 id="weight-srd1gt2tz">Weight</h2>Doesn't have much of an effect at low weights. 1.0 is a starting point, 1.2 was used for most sample images.<h2 id="tags-cosnhkyu9">Tags</h2>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:<pre><code>hstyle, hs03, hs08
hstyle, hs03, hs08, hs17, hs49
hstyle, hs00, hs04, hs11, hs19, hs37
hstyle, hs05, hs06, hs09</code></pre>Other useful tags:<ul><li>anime coloring : Style reinforcement</li><li>vum : Hashed token I found that produces a vintage style, some might enjoy</li></ul>Negative tags:<ul><li>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.</li></ul><h2 id="dataset-b21nw69kk">Dataset</h2>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 &amp; outputs, not the full dataset ofc.

Files with SHA256: aba16849dcb994ccf687870d02101e7298ecffe77be67b103f73909acc3aab43