Egg Fusion - LoRa Merge

512dim lora および chilloutmix_NiPrunedFp32Fix + deliberate_v2 をベースに完全に刷新されたチェックポイント。

トレーニングデータ: 512 DIM LORA
/model/41893/lora-eggeaster-fusion

以下は、モデル上に lora を適用した際の様子:

lora 重量: TEnc Weight 0.2 UNet Weight 1

512dim lora を chillout および deliberate にマージ / --ratios 0.99 を Egg Fusion lora へ

C は CHillout / d は deliberate

C と D を以下でマージ:

Base_alpha=0.53
Weight_values=0,0.157576195987654,0.28491512345679,0.384765625,0.459876543209877,0.512996720679012,0.546875,0.564260223765432,0.567901234567901,0.560546875,0.544945987654321,0.523847415123457,0.5,0.476152584876543,0.455054012345679,0.439453125,0.432098765432099,0.435739776234568,0.453125,0.487003279320987,0.540123456790124,0.615234375,0.71508487654321,0.842423804012347,1

および
1,0.842423804012346,0.71508487654321,0.615234375,0.540123456790123,0.487003279320988,0.453125,0.435739776234568,0.432098765432099,0.439453125,0.455054012345679,0.476152584876543,0.5,0.523847415123457,0.544945987654321,0.560546875,0.567901234567901,0.564260223765432,0.546875,0.512996720679013,0.459876543209876,0.384765625,0.28491512345679,0.157576195987653,0

両方の結果をマージ

Weight_values=
1,0.842423804012346,0.71508487654321,0.615234375,0.540123456790123,0.487003279320988,0.453125,0.435739776234568,0.432098765432099,0.439453125,0.455054012345679,0.476152584876543,0.5,0.523847415123457,0.544945987654321,0.560546875,0.567901234567901,0.564260223765432,0.546875,0.512996720679013,0.459876543209876,0.384765625,0.28491512345679,0.157576195987653,0

そして、ご覧の通りです。
512 dim で学習された lora を搭載した Egg_fusion チェックポイント

お楽しみください 🤟 🥃