<h3 id="concept">Concept</h3>A helper model to provide areola and nipple variety, instead of the default round pink nipples you get out of the box with WAN. Works with both T2V and I2V. One thing to be careful of is that many mixes throw in content that will force default areolas. This has been largely tested and trained on vanilla WAN 2.2. Definitely the bumpy keyword with I2V can produce some horrifying results with the wrong checkpoints.<h3 id="usage">Usage</h3>The models aims to provide a 5 different parameters to change areolas and nipples; color, texture, size, shape and length. Prompts can combine these into single statements or a few different statements. The aim was to provide the most flexibility but sometimes one type of aspect is hard to find combos of(for example large areolas on women with flat chests).Base prompt is:<pre><code>She has ___ areolas</code></pre><pre><code>She has ___ nipples.</code></pre>You can join keywords together but i find 3 is about the max a single statement can be get effective. For example 'She has dark bumpy oval areolas' should work but may not be as effective as each one being written individually.Examples would be:Color:<pre><code>She has dark areolas.</code></pre><pre><code>She has pale areolas.</code></pre><pre><code>She has pink areolas.</code></pre>Texture:<pre><code>She has puffy areolas.</code></pre><pre><code>She has smooth areolas.</code></pre><pre><code>She has bumpy areolas.</code></pre>Size:<pre><code>She has large areolas.</code></pre><pre><code>She has small areolas.</code></pre>Shape:<pre><code>She has oval areolas.</code></pre><pre><code>She has round areolas.</code></pre><pre><code>She has tuberous breasts.</code></pre>Length:<pre><code>She has long nipples.</code></pre><pre><code>She has short nipples.</code></pre><pre><code>She has flat nipples.</code></pre><pre><code>She has inverted nipples.</code></pre>Weight of 1.0 for both high and low should work well. Sampler wise most should work but i'd avoid unipc as it does some weird things. Stick with Euler/LCM/Res.<h3 id="training">Training</h3>This is my first time training a T2V with the aim of using it on both model types, so I'd be interested in how people find it. I'm not sure if training with no videos will be a massively negative thing, i guess we'll see!Trained at 0.6 MP for high and 1 MP for low. Batch of 12 on high and 6 on low. Dataset used was 187 images only, roughly 15 per concept, no videos used at all. 40 Epochs on high and 58 on low. Captioning was automated using JoyCaption. Trained on diffusion pipe on 3x3090s and 1x5060ti.

Files with SHA256: 52a5d0d000ae30b6234f4d66ced322a5ac74cf710480a0d1b0d3cd1220d870c1