<h1 id="heading-1145">VanillaCore</h1><h2 id="heading-1146">(Official change in name from HentaiCore)</h2><h3 id="heading-1147">Complete rework of the original HentaiCore on training and dataset. Trained on 70 sex position (concepts) with +40 images each one.</h3><h3 id="heading-1148">Learn to prompt at the end!!</h3>Don't use VAE!!!<h3 id="heading-2">If you want to support me, can buy me a coffee. Thanks!!</h3><a target="_blank" rel="ugc" href="https://www.buymeacoffee.com/Norian01">https://www.buymeacoffee.com/Norian01</a>DreamCorev3 Changelog:<ul><li>Absolute rework in the dataset and was trained for 10 hours on a A100</li><li>Works perfectly with character loras trained on NAI</li><li>The version with higher consistency and simpler prompting</li><li>Taining refined for over +30 experiments before final model</li><li>Expanded concepts from 40 to 70 while increasing number of images in each one</li><li>Trained a little in landscapes, mans, and a slight different style through regularization</li></ul>Specifications of training:3100 images~2 repeats24 epochs10 batchBase Model: Hassaku (Hentai Model)<h3 id="heading-2">This model has a very easy prompting. It normally follows the following structure:</h3><h3 id="heading-1149">Activation tags + 1girl, pussy, 1boy, penis, vaginal, sex</h3>You can add to the second part traits, costumes, characteristics of the characters. But you need to include those tags above almost allways (less pussy, penis, vaginal, or breasts when none of them are seen) and you need to add things like breasts in paizuri, all logical.Follow my sample images for if you are missing something.Activation Words (listen, a (+) means that the following tags are used to get a different angle of the position, they need to be aside of the activation tag separated by comma, and you can only use one angle at a time)(Ex: doggystyle + pov + from side and i choose doggystyle, from side)USE CLIP SKIP 2!!!Dont use VAE, it will saturate the image in this model!!!<ul><li>amazon position</li><li>cowgirl position + from behind + pov</li><li>reverse cowgirl position + from behind + from side + pov + pov, standing</li><li>squatting cowgirl position</li><li>doggystyle + pov + from side + from below + standing + standing, from side + standing, pov + fellatio + top-down bottom-up</li><li>jack-o challenge + from behind</li><li>lying on person + on back</li><li>lying split, bed</li><li>standing split</li><li>standing missionary</li><li>mating press + deep penetration + from side</li><li>missionary + pov + legs up, pov</li><li>piledrive</li><li>prone bone + from side + pov</li><li>suspended congress + from side</li><li>reverse suspended congress</li><li>upright straddle + from side + pov</li><li>reverse upright straddle</li><li>spooning position</li><li>breast grab, pov + cowgirl position + missionary</li><li>leaning forward + standing + cowgirl position + squatting cowgirl position</li><li>guided breast grab + missionary + cowgirl position + squatting cowgirl position</li><li>fellatio + pov + from side + behind perspective + cheek bulge + 2girls + 3girls + from side, squatting, full body + from side, full body</li><li>licking penis + from side</li><li>paizuri</li><li>female masturbation + (prompt a pose with regular tags)</li><li>fingering + (prompt a pose with regular tags)</li><li>aftersex + (desired position, doesnt work with all)</li><li>imminent penetration + (desired position, doesnt work with all)</li><li>x-ray cervix + internal cumshot + (desired position, doesnt work with all)</li><li>x-ray mouth, fellatio</li><li>cross-section + internal cumshot + (desired position, doesnt work with all)</li><li>arm grab + (desired position, doesnt work with all)</li><li>grabbing own breast + (desired position, doesnt work with all)</li><li>guided penetration + (desired position, doesnt work with all)</li><li>holding hands + (desired position, doesnt work with all)</li></ul>

VanillaCore (hentai model)__70 positions trained

Changes in this version:<ul><li>Version 2 has a more strong style, but also more consistency in the positions.</li><li>Using add difference merge to have a more concentrated effect of the training, instead of sum in the version 1.</li><li>Using for merge Hassaku instead of Grapefruit, it`s merged because the style of trained model is very strong.</li><li>Also merged BetterHands Locon. </li></ul>

Files with SHA256: b1a21dfded5825fd6dbe33d5a8a98d9bc92653b5aa3416150d1f671f7112b46b