license: creativeml-openrail-m
real_model_L
recipe only for personal use. (not for commercial, because of license)
This model "file" is not public anymore, I try to change some asset and weight, I'll share next model.
photorealistic checkpoint for sd1.5, model merge example.
recipe for supermerger:
F is LoRA merge to checkpoint. D,G,H,I,L are using MBW and weight sum. J is using sum twice. other is using weight sum.
- A = cityedgemixV1_v125 x 0.5 + kisaragiMix_v22 x 0.5
- B = majicmixRealistic_v6 x 0.5 + shampooMix_v4 x 0.5
- C = A x 0.5 + B x 0.5
- D = fantasticmix_v65 x (1-alpha) + dreamshaper_631BakedVae x alpha (0.4,0.35,0.4,0.45,0.45,0.3,0.3,1.0,0.0,0.0,0.0,0.0,0.0,0.0,1.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0)
- E = C x 0.8 + D x 0.2
- F = E + flat2:-0.7 (lora marge)
- G = F x (1-alpha) + calicomixreal_v20 x alpha (1.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0)
- H = F x (1-alpha) + tauronHybridMix_tauHybridRealV21 x alpha (1.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0)
- I = F x (1-alpha) + xxmix9realistic_v30 x alpha (1.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0)
- J = (G x 0.6 + H x 0.4) x 0.6 + I x 0.4
- K = J x 0.439 + savMIX_xl x 0.561
- L = K x (1-alpha) + kencanmix_v16 x alpha (0.0,0.0,0.0,0.0,0.0,0.0,0.5,0.0,0.5,0.5,0.5,0.0,0.0,0.0,0.0,0.0,0.0,0.15,0.25,0.35,0.5,0.0,0.0,0.0,0.0,0.0)
Thanks to the makers of those wonderful models and LoRAs!
The model file is not available, but you can try to merge the models. ...welcome to model merge swamp! (ようこそモデルマージ沼へ)
I'm not a native English speaker,(I'm tired,,,) so I wrote follow descriptions in japanese.
結構感覚的に作ってたんだなぁと思う作成時の記録を下記に書きます
A～Cはいい感じにかわいいアジア系の女の子の完全な写真が出ると思われるモデルを均等にマージ(これを基本系とするため。ここは正直雑に混ぜたので今後の改善ポイントかもしれない)
Dでちょっとだけ2Dの入ってるdreamshaperの形状や構造を取り入れたかったのとfantasticはしっかり写真で反応良かったので混ぜる
Eでここまで作ったものを平均化
Fで詳細化をかけておく(-1はやりすぎかなと思ったので-0.7にした)
GーIでTE変更(とりあえずマージ候補として選定していた中で特に2次元キャラのLoRA(主に衣装)に正確に反応してくれる3Dの厳選したモデルを使った)
Jで比率見ながらMIX(tauが他への影響が強かったのでちょっと弱めた)
Kでsavを何となく取り入れる(出力したらかなり良い画像が出てきていたので取り入れたかった)
Lでkencanmixの顔層を取り入れて(衣装に影響が出るのでOUT側は若干抑制している。これ以上OUT側を増減すると衣装と顔の出力が微妙になるのでギリギリこの値)
そのあと色々混ぜてみたもののなかなかうまくいかず……結局これが一番良かったのでこれにしました。
特定のseedと特定のLoRAの組み合わせでしかテストしていないです(気晴らしで他のLoRAが3次元化することは確認しています…一部は目が大きすぎたりするので若干LoRAの比重を下げたりはしましたが…)
そのためあまりしっかり出ないLoRAもあるかもしれませんが、そんな時はそのLoRAがしっかりと出るモデルをマージして作ってみるのも一興かもしれません(みんなでモデルマージ沼に浸かろう)