Sign In

Katheryne (Genshin Impact NPC) キャサリン・凯瑟琳 (原神)

219
1.1k
15
Verified:
SafeTensor
Type
LoRA
Stats
1,114
Reviews
Published
May 10, 2023
Base Model
SD 1.5
Training
Steps: 1,590
Epochs: 10
Hash
AutoV2
8D31E043AD
default creator card background decoration
Obliqeray's Avatar
Obliqeray

Made from 32 screenshots of her 3d model (3 processed through IMG2IMG to make the face nicer) and 19 images made using an older version of this LoRA and various checkpoint models, to try and offset the 3d model look. Not happy with the result, but I'm busy, so I dunno if I'll try redoing it later or not.

参考データは3dモデルのスクショ32枚(顔を綺麗にするため、3枚はIMG2IMGで処理した)と、3dモデルのスタイルから離れるために、このLoRAの古い版といろんな別スタイルで作った19枚の絵でした。この結果に満足してないけど、忙しいからいつかやり直すかわかりません。

Recommended prompts (おすすめタグ):

short hair, medium hair, jewelry, gem, green and white dress, off-shoulder, bow, white pantyhose, vertical-striped pantyhose, puffy sleeves, blunt bangs, detached collar, maid headdress, red bow, green collar, green shoes, corset, blue eyes, black hair

Recommended weight (おすすめウエイト):

0.7-1.0

Even at these weights, the details tend to be slightly off depending on the checkpoint model used. You'll also need to use the recommended prompts.

このウエイトでも、使ってるチェックポイントモデルによって詳細がよく少しずれてます。おすすめタグも使う必要です。

From my experience, LoRAs seem to work by overwriting what the AI originally intended to draw, based on your prompts. So the more complex a character is, the more you'll need to explain the character's basic appearance. For example, if the AI intended to draw a tight swimsuit, and your LoRA is a big dress, it'll probably make a swimsuit that looks like your LoRA dress (ie. with the skirt missing).

今までの個人体験ですが、LoRAは使ってるタグによってAIが元々描くつもりのキャラクタを書き換えてるっぽいです。ということは、キャラのデザインがもっと複雑場合、その外見の基本をもっとはっきり説明する必要だと思います。例:元々キツイ水着を描くつもりで、使ってるLoRAは大きなドレス。結果は多分LoRAのドレスに似てる水着(スカートなし)の絵になります。

Problems (問題点):

  • Patterns and minor details are often slightly off.

    パターンと細かい詳細がよく少しずれてます。

  • At 0.8-1.0 the style of the face will be less influenced by the checkpoint model.

    0.8-1.0の場合、顔にチェックポイントモデルからの影響が減ります。

Models used for screenshots (スクショに使ったモデル):

  • Counterfeit v3.0

  • Based65 Final Mix

  • Tsubaki v1.0

  • CardosAnime v1.0

  • ExpMixLine v2

  • Meina Mix v9

  • SweetMix v1.6/1.7

  • CoffeeMix v2.0

  • Anything v4.5 (with Hoyo Artstyle LoRA v2)