<p>The voicemod protogen, trained using the dreambooth extension for Automatic1111 webui. it was trained on only 5 images of the voicemod protogen, combined with 10 images of generic protogens, all simply tagged to get the model to understand more concepts that can apply to protogens.</p><p></p><p>I had previously attempted it with just the 5 images I have of the character, this caused it to lack knowledge of features that are common with protogens, and made it home towards generating images that look like copies of the training data. That's why I included the 10 extra images. And it worked great!</p><p></p><p>This is the first LoRA I am publishing, after playing around with LoRAs for a bit I think this one is good enough to publish.</p><p></p><h2>training</h2><p>The LoRA was trained using the dreambooth extension for Automatic1111 webui. I used 15 images of protogens, of which 5 were of the voicemod protogen specifically. I used deepbooru to caption the images, then edited the captions to use spaces instead of underscores, and added more specific information, removed incorrect tags.</p><p>I generated 75 classifier images. And just let the LoRA train for 2 hours.</p><p>For the base model I trained it on was a 50/50 merge of AbyssOrangeMix3 and Anything v4.5. However, the LoRA should work fine on most models.</p><p></p><h2>how to use:</h2><p>main trigger word: "voicemod protogen"</p><p>additional trigger words: "striped thighhighs", "striped arm warmers", "[color] visor", "multicolored fur"/"multicolored hair", "[facial expression] expression"</p><p>recommended to have in negative prompt: "text", "speech bubble", "watermark"</p><h2>known limitations:</h2><ul><li><p>unable to generate 3d protogens, the training data was almost exclusively 2d and it seems it has associated the character to be in 2d, using it on a model that usually generates 3d content will most likely cause it to generate the character in 2d.</p></li><li><p>changing the visor color will change the main theme color of the protogen as well, and it might be inconsistent too.</p></li></ul><h2>note</h2><p>the model is capable of generating different protogens, however you will still need to specify "voicemod protogen", just putting "protogen" often works but has issues with the visor. instead, specify "voicemod protoen" and specify the other details you want changed.</p>

Files with SHA256: 8d34eafcfb6c76649ff55a70f028a7e6d359a6697e9cbc00ad464565bc18abdf