Sign In

【Character】宇沢レイサ | uzawa reisa (Blue Archive)

58

338

0

Verified:

SafeTensor

Type

LoRA

Stats

338

0

Reviews

Published

Sep 10, 2023

Base Model

Other

Training

Steps: 2,976
Epochs: 12

Usage Tips

Clip Skip: 2

Trigger Words

uzawa reisa
1girl
low twintails

Hash

AutoV2
5DF2489CE2
default creator card background decoration
Morinn's Avatar

Morinn

一些训练关键参数:

先验损失权重 (prior_loss_weight) = 1

分辨率 (resolution) = "832,832" , batch_size = 3

unet_lr = 1 , text_encoder_lr = 1 , lr_scheduler = "cosine_with_restarts" ,

优化器 (optimizer_type) = "Prodigy" , d_coef=2

最小信噪比伽马 (min_snr_gamma) = 3

network_dim = 64 , network_alpha = 32

最大正则化范数 (scale_weight_norms) = 1

分层训练学习率权重设置:

编码层 (down_lr_weight) = "1,1,1,1,1,1,1,1,1,1,1,1",

中间层 (mid_lr_weight) = "1"

解码层 (up_lr_weight) = "1,1,1,1,1,1,1,1,0,0,0,0"

置0阈值 (block_lr_zero_threshold) = "0.1"

训练图片总计124张,12个epoch,总计2976步

分别在Anything V5、AbyssOrangeMix、CreationMix、GhostMix、DiaMix、QteaMix、DonutHoleMix、CounterfeitV30、CuteYukiMix 和 DreamShaper 上进行了测试。其中 Counterfeit、CreationMix 、DiaMix 和 DonutHoleMix 使用了 vae-ft-mse-840000-ema-pruned。可以猜一猜每张样图都对应哪一个底模。

部分提示词在不同模型上,存在一定的欠拟合或过拟合。比如出原设的时候,在Anything上会多出一双猫耳;在提示词数量较少的情况下,多个模型均出现了提示词中不存在的双肩包。

由于是试验性作品,且目前泛化性还算正常,因此暂时不考虑优化(因为涉及到简化训练集的tag,工作量太大)

建议使用的prompts:

①人物:“1girl,solo,uzawa reisa,low twintails”

②校服:“school uniform,sneakers,backpack,serafuku,neckerchief,jacket,striped socks,pleated skirt,white sailor collar”

如果喜欢,请给个好评