generated at
LyCORIS
GitHub : https://github.com/KohakuBlueleaf/LyCORISLora beYond Conventional methods, Other Rank adaptation Implementations for Stable diffusion.
Stable Diffusionやその他に対してパラメータ効率的なfine-tuningを行うための異なるアルゴリズムを実装するためのプロジェクト
LoConブランチからのスタート

実装したもの
LyCoris
FullLoRALoHaLoKr low factorLoKr high factor
Fidelity
Flexibility
Diversity
Size
Training Speed Linear
Training Speed Conv
★ > ◉ > ● > ▲ means better and smaller size is better
LoHa : LoRA with Hadamard Product representation
左の図が2R個のランク1行列(縦ベクトルと横ベクトルの積)の総和になるのに対し、右の図はRの2乗個のランク1行列の総和になるので同じパラメータ数でランクを大きくできるらしい。
This algo produce very tiny file(about 200~300KB)
ちっちゃwogikaze
2.1.0
2.2.0
QLycCORIS with LoHa/LoKr/OFT/BOFT
bitsandbytes8bit/4bit Linear layer

LyCORISモデルといっているけど中身はLoCoNとかLoHAであって、"LyCORIS"という技術で作られたモデルはないっていう認識で合ってるのかな?nomadoor

リコリス?wogikaze
GitHubのトップがこれなので、開発者のセンスが爆発してますねクォレハ…nomadoor