This LoRA model is similar to my HyperAss model but focuses more on thick thighs and hips. Great for that pear shaped look. You can always increase the LoRA strength for a bigger effect.If you don't want crazy large sizes, just keep the LoRA strength at 1. It does in-between sizes just fine.<h3 id="heading-6">v2 Changelog 2023/06/13:</h3>Damn, its been almost 5 months o.o<ul><li>Improved the overall quality. Less overcooked results at high strength.</li><li>The shape should be even more pear like now, smaller top and larger bottom in general.</li><li>Added ass size tags, similar to my other models. See tag data attached with model in the downloads section for details. A few tags are pretty important.</li><li>Uses Kohya's LoCon LoRA but does not require any additional extensions to run.</li><li>Doubled the dataset size.</li><li>Added tag "from front" to assist with front facing shots, and a few others in the tag docs</li></ul>Notes:I used this to train my image tagging classifiers for ass sizes <a target="_blank" rel="ugc" href="https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification">https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification</a>

hyper bottom heavy

Similar to my last released hyperbreasts model, this one should be less damaging to the overall style while also making it easier to achieve large sizes. Also I improved the accuracy of the tags, and doubled the dataset size to ~400 images.Increasing LoRA strength is actually useful for achieving larger sizes in combination with the right size tags.Training Details:<ul><li>~400 images</li><li>160 epocs</li></ul><ul><li>learning rate 2e-4</li></ul><ul><li>text encoder LR 1e-4</li><li>base model Av3</li><li>clip skip 2</li><li>random flip</li><li>tag drop chance 0.15</li><li>network dropout 0.25</li><li>bucketing at 768</li><li>dim 32</li><li>alpha 16</li><li>225 tokens</li><li>cosine with restarts</li><li>training with tags, tags attached next to model download<ul><li>use the new weighted captions + dropout in Kohya that way more important tags were trained at a higher weight (weight of:2).</li></ul></li><li>Kohya LoRA LoCon, does not require any additional extensions to use</li></ul>

Files with SHA256: 4819200ac502ee57736b0f6e807e451e19c1fd534847caba9b427a35bd75b8a8