A character LoRA for Navia from Genshin Impact. Made from images from the Fontaine Overture Teaser PV.Just a warning that I can't guarantee the quality of the LoRA since I'm new to making LoRAs ,and I wanted to see how well I could with a very limited dataset. Due to using only a few images, it's probably overfit in some weird places and you will probably experience style bleeding with the LoRA. Regularization images were added to help with the flexibility of the LoRA but you can encounter underfitting if you exclude too many characteristics or use non-frontal viewpoints.<h1 id="heading-4766">Usage</h1>Recommend Weight: 0.6 ~ 0.8 (You can try using the LoRA block weight extension as well. I don't have a specific recommendation despite using OUTD myself).Trigger Words: <pre><code>navia,naviagenshin,1girl</code></pre>General Appearance:<pre><code> blonde hair,blue eyes,swept bangs,ringlets,very_long_hair,drill_hair</code></pre>Outfit: <pre><code>detached_collar,detached_sleeves,necklace,pearl_necklace,elbow gloves,dress,thigh_boots,high_heels</code></pre>Tips for generating the Hat:<pre><code> rose,hat,black_headwear,hat_feather, tricorne, witch hat</code></pre><ul><li>Difficult to get the hat working correctly at times</li><li>(Not sure why witch hat occasionally helps)</li></ul><h2 id="heading-4767">Overfitting</h2><ul><li>(thigh_boots, and high_heels have heavy bias to full_body poses)</li><li>Her grin is overfit so you need to use emphasis for other facial expressions</li></ul><h1 id="heading-4768">Dataset</h1><ul><li>5 Images from the PV cropped into 17 parts.</li><li>56 Regularization images</li></ul>Main limitation seems to the lack of higher resolution images so that SD doesn't hallucinate. The regularization also causes the eyeshape to underfit so the LoRA will prefer the checkpoint's styling for eyes.

[Genshin Impact] Navia

initial version<ul><li>dataset from PV</li><li>regularization from isabella lora</li></ul>

Files with SHA256: 1c4b2e3527fac5d5aca17c420e291e06c8d8e31b0711a4336442972b7e7f2704