A concept LoRA designed to bring stability to <a rel="ugc" href="https://danbooru.donmai.us/posts?tags=spoken_symbol">spoken symbols</a> (i.e. symbols in a speech bubble) for anime-style image generations using <a rel="ugc" href="https://civarchive.com/models/257749/pony-diffusion-v6-xl">Pony</a>. Out-of-the-box Pony tends to produce incoherent doodles when given most spoken symbol prompts (with the exception of <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/posts?tags=spoken_heart">spoken heart</a>). This LoRA took some rigorous testing and several training attempts to get a good balance between new and modified concepts whilst having a negligible impact on style. The LoRA was trained with 11 spoken symbols, the prompts for which are listed below. The probability of generating an accurate symbol when prompted is also included:<ol><li>Spoken heart, (≈100% reliability)</li><li>Spoken star, (≈100%)</li><li>Spoken blush, (90%)</li><li>Spoken sweatdrop, (90%)</li><li>Spoken x, (90%)</li><li>Spoken venus symbol, (85%)</li><li>Spoken anger vein, (80%)</li><li>Spoken musical note, (75%)</li><li>Spoken mars symbol, (70%)</li><li>Spoken o, (60%)</li><li>Spoken skull, (55%)</li></ol>Future version will improve these probabilities and include more spoken symbols.

Files with SHA256: 3b00c44298b87a86035576e4decd34c9cc8a33718f199c8ee5ef7d1147132dee