This LoRA facilitates imitating the style of Yadokari (also spelled Yadkari, or やどかり), an illustrator best known for their work on character art and illustrations for the browser game and media franchise Kantai Collection.<h2>Prompting</h2>No activation tokens are required to use this LoRA model. If the style of the eyes ends up looking a bit lifeless, try adding/strengthening "finely detailed iris".<h3>v6 models</h3>For v6 models I recommend using weights between 0.3-0.9, depending on how stylized you want it to be.Character feature recall for Yadokari's characters can be hit-and-miss, due to low frequency of each individual character in the dataset. Regardless, prompting for features will typically allow proper recall and it has learned some aspects. This is (partially) addressed by the v8 model, but at the cost of ease-of-use and versatility.<h3>v8 models</h3>For v8 models I recommend using weights between 0.2-0.6, depending on how stylized you want it to be.Character feature recall is improved from v6 models due to better tagging &amp; higher text encoder learn rate, but this also means images more easily start to look like the training set. Most notably, simpler backgrounds &amp; rigging appear more easily.<h1>Dataset</h1>About 92 images of artworks by Yadokari were scraped from Danbooru, consisting of mostly KanColle but there are also fanarts of other IPs and fully original works. In addition, the training set also included 50 unique colored 2D works scraped from Yadokari's Pixiv and Twitter. Some artworks were rotated to be more upright orientation if the image did not contain very clear elements of gravity that would seem unnatural if rotated. Duplicates, or closely related images (for instance, intermediate drafts of later artworks, crops of other artworks) were pruned.Tagging was done via wd14-swinv2-v2 with threshold 0.15.

Yadokari Style LoRA

If you only want to imitate the artstyle and don't particularly care about portraying a Yadokari ship girl, I highly recommend using v6 instead.Trained on AOM2-nutmegmixGav2. Compared to v6-AOM2-nnmGav2, the model now more closely imitates details of the underlying dataset. This means it will both reproduce character traits more accurately, but also it will tend to crush background details (most the training set is simple background) and can more easily ignore prompts (see the attached UNet/TEnc grid).Example ImagesMade using AOM2-nutmegmixGav2 txt2img LoRA weight of about 0.5, then upscaling with Ultimate SD Upscale and no further processing. Image grids are txt2img only.Negative prompts utilizes bad-artist and bad-artist-anime embeddings from <a target="_blank" rel="ugc" href="https://huggingface.co/NiXXerHATTER59/bad-artist">https://huggingface.co/NiXXerHATTER59/bad-artist</a> as well as the bad-hands-5 embedding from <a target="_blank" rel="ugc" href="https://huggingface.co/yesyeahvh/bad-hands-5/tree/main">https://huggingface.co/yesyeahvh/bad-hands-5/tree/main</a>Hiryuu image uses a non-public LoRA.Training Parameters"net_dim": 128,"alpha": 128.0,"scheduler": "cosine_with_restarts","cosine_restarts": 3,"warmup_lr_ratio": 0,"learning_rate": 0.0001,"text_encoder_lr": 5e-05,"unet_lr": 0.0001,"shuffle_captions": true,"keep_tokens": 5,"train_resolution": 768,"lora_model_for_resume": null,"unet_only": false,"text_only": false,"vae": null,

Files with SHA256: c195a160779fb9742892404efe9ddc78ced4a9c4131b1376a7d3772d39a7d2f9