deepfacelab中文网

 找回密码
 立即注册(仅限QQ邮箱)
查看: 1775|回复: 10

速写画风——最难的抽象风格类Lora训练日记

[复制链接]

1

主题

2

帖子

190

积分

高级丹童

Rank: 2

积分
190
 楼主| 发表于 2024-3-26 18:53:34 | 显示全部楼层 |阅读模式
本帖最后由 fuaneng 于 2024-3-26 18:58 编辑

声明:改帖子是我已经在知乎发布的文章,在此发布以供大家学习借鉴之用,并非盗帖!

一、数据集处理
这里就不上传我训练的数据集了,如果大家需要也可以留言找我要。

二、打标签以及固定触发词
调取 Lora 的Prompt:sketch,portrait,monochrome,greyscale,
序号分类触发词
1老男人elderly man
2老妇人elderly woman
3中年男人middle aged man
4中年妇女middle aged woman
5年轻男人boy
6年轻女孩girl
7小男孩little boy
8小女孩little girl

三、Kohya_ss 训练脚本文件结构

3.1 训练集文件夹结构:


3.2 正则化文件夹结构:


四、训练参数

4.1 详细参数
  
"LoRA_type": "Standard",  // LoRA类型为标准  "LyCORIS_preset": "full",  // LyCORIS预设为完整  "adaptive_noise_scale": 0,  // 自适应噪声比例为0  "bucket_no_upscale": true,  // 桶无放大_启用√  "bucket_reso_steps": 64,  // 桶分辨率步数为64  "cache_latents": true,  // 缓存潜在表示_启用√  "caption_dropout_every_n_epochs": 0.0,  // 每n个时期的标题丢弃率为0.0  "caption_dropout_rate": 0,  // 标题丢弃率为0  "clip_skip": 2,  // 剪辑跳过为2  "conv_alpha": 1,  // 卷积alpha为1  "conv_dim": 1,  // 卷积维度为1  "debiased_estimation_loss": false,  // 无偏估计损失_不启用❌  "decompose_both": false,  // 分解_不启用❌  "dim_from_weights": false,  // 从权重中获取维度_不启用❌  "enable_bucket": true,  // 启用桶_启用√  "epoch": 10,  // 训练纪元为30  "factor": -1,  // 因子为-1  "flip_aug": false,  // 翻转增强_不启用❌  "fp8_base": false,  // FP8基础_不启用❌  "full_bf16": false,  // 完全BF16_不启用❌  "full_fp16": false,  // 完全FP16_不启用❌  "gradient_accumulation_steps": "1",  // 梯度累积步数为"1"  "learning_rate": 0.0001,  // 学习率为0.0001  "logging_dir": "D:/fyn_ai/kohya_ss/training/log",  // 日志目录为"D:/fyn_ai/kohya_ss/training/log"  "lr_scheduler": "cosine_with_restarts",  // 学习率调度器为余弦重启  "lr_warmup": 5,  // 学习率预热为5,即为5%  "max_bucket_reso": 2048,  // 最大桶分辨率为2048  "max_resolution": "960,960",  // 最大分辨率为"960,960"  "max_timestep": 1000,  // 最大时间步长为1000  "max_token_length": "75",  // 最大令牌长度为"75"  "min_bucket_reso": 256,  // 最小桶分辨率为256  "mixed_precision": "fp16",  // 混合精度为FP16  "model_list": "custom",  // 模型列表为自定义  "multires_noise_discount": 0,  // 多分辨率噪声折扣为0  "multires_noise_iterations": 0,  // 多分辨率噪声迭代次数为0  "network_alpha": 1,  // 网络alpha为1  "network_dim": 128,  // 网络维度为128  "network_dropout": 0,  // 网络丢弃为0  "noise_offset": 0.05,  // 噪声偏移为0.05  "noise_offset_type": "Original",  // 噪声偏移类型为原始  "num_cpu_threads_per_process": 2,  // 每个进程的CPU线程数为2  "num_machines": 1,  // 机器数为1  "num_processes": 1,  // 进程数为1  "optimizer": "AdamW8bit",  // 优化器为AdamW8bit  "output_dir": "D:/fyn_ai/kohya_ss/output",  // 输出目录为"D:/fyn_ai/kohya_ss/output"  "output_name": "lbc_lbc_Sketch style",  // 输出名称为lbc_lbc_Sketch style  "pretrained_model_name_or_path": "D:/sd-webui-aki/sd-webui-aki-v4.1/models/Stable-diffusion/realisticVisionV60B1_v51VAE.safetensors",  // 预训练模型路径为"D:/sd-webui-aki/sd-webui-aki-v4.1/models/Stable-diffusion/realisticVisionV60B1_v51VAE.safetensors"  "seed": "31337",  // 种子为"31337"  "text_encoder_lr": 1e-05,  // 文本编码器学习率为1e-05,即为0.00001  "train_batch_size": 3,  // 训练批大小为3  "train_data_dir": "D:/fyn_ai/kohya_ss/training/img",  // 训练数据目录为"D:/fyn_ai/kohya_ss/training/img"  "train_on_input": true,  // 训练输入为真  "unet_lr": 0.0001,  // UNet学习率为0.0001  "xformers": "xformers"  // 变换器为xformers

4.2 Konya_ss 训练脚本界面参数




4.3 基础参数记录及训练过程分析


本次训练了30纪元,下方我直接用曲线图表现Lora训练时的拟合情况


通过上面表格和曲线图,我们可以从直观了解到本次训练模型的基本信息和训练过程的拟合情况,共计训练迭代了30个版本,从拟合曲线中我们可以得知,本次训练Lora的拟合值(loss值)是基于0.108这个数值上下波动的。
由于我选择的调度程序是cosine_with_restarta(可重启的余弦函数),所以它是具有波动性的。观察曲线图可以得知大约有9次拟合情况(理想上来说可以得到10个不错的风格模型),不过到这一步我们也只能是猜测,并不能确定处于这个loss值的Lora一定符合我们的要求。
不过,从训练过程中loss值表现看来,本次的lora模型在训练中的表现是非常趋稳定的。

五、测试

直接调用 stable diffusion webui 的 xyz plot 脚本进行测试,更直观地查看模型拟合情况。

NUM,000001,000002,000003,000004,000005,000006,000007,000008,000009,000010,000011,000012,000013,000014,000015,000016,000017,000018,000019,000020,000021,000022,000033,000024,000025,000026,000027,000028,000029,000030
STRENGTH,0.4,0.6,0.8,1


使用的正、负面提示词
sketch,a elderly woman,portrait,monochrome,greyscale,<lora:lbc_lbc_Sketch style-NUM:STRENGTH>,
extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),


使用的底模
写实底模:realisticVisionV51_v51VAE.safetensors [ef76aa2332]
二次元底模:AnythingV5V3_v5PrtRE.safetensors [7f96a1a9ca]




基于以上两组不同底模(二次元底模和写实底模)的测试,我初步选了000018、000022、000026、000029这四个模型。
由于上述的常量写错了一个(将 000023 写成了 000033 ),所以在测试图中有一个模型没有被测试到,不过并不不影响总的结果。
当然我后后也会将这个没测试过的模型放上来!

进一步对挑选出来的5个Lora模型进行测试:
针对之前选出000018、000022、000023、000026、000029Lora模型,再次进行了拟合度测试,选出最为拟合的模型。
parameters(本次测试的指定人物为老年男性):
sketch,elderly man,portrait,monochrome,greyscale,<lora:lbc_lbc_Sketch style-NUM:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1608443307, Size: 512x512, ENSD: 31337, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "NUM,000018,000022,000023,000026,000029", Y Type: Prompt S/R, Y Values: "STRENGTH,0.4,0.6,0.8,1", Version: v1.8.0

使用底模为写实模型(Model: realisticVisionV60B1_v51VAE,)表现效果如下:


使用底模是二次元模型(AnythingV5V3_v5PrtRE),测试效果如下:


使用泛化性较低的亚洲人像写实类底模测试结果如下:

parameters(中年男子)
sketch,monochrome,greyscale,middle aged man,mid_shot,full_shot,<lora:lbc_lbc_Sketch style-NUM:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1986768555, Size: 512x768,
Model: None南-亚洲男生摄影_V2, Clip skip: 2, ENSD: 31337


基于训练底模测试(中年女性)表现效果
parameters
sketch,monochrome,greyscale,middle aged woman,mid_shot,full_shot,<lora:lbc_lbc_Sketch style-NUM:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3768944837, Size: 512x768,
Model: realisticVisionV60B1_v51VAE, Clip skip: 2, ENSD: 31337,


基于训练底模测试(boy男孩)表现效果
parameters
sketch,monochrome,greyscale,boy,portrait,<lora:lbc_lbc_Sketch style-NUM:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 2244857732, Size: 512x768, Model: realisticVisionV51_v51VAE, ENSD: 31337,


基于训练底模测试(girl女孩)表现效果
parameters
sketch,monochrome,greyscale,girl,portrait,white_background,<lora:lbc_lbc_Sketch style-NUM:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 400776507, Size: 512x768,
Model: realisticVisionV60B1_v51VAE, ENSD: 31337,


综上,我个人觉得模型000023是比较不错的选择,当然其他模型我也会在附件中上传,下面我会对模型000023进行泛化性和拟合度测试。
parameters
sketch,monochrome,greyscale,girl,portrait,white_background,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 499890659, Size: 512x512, Model hash: 7f96a1a9ca, Model: AnythingV5_v5PrtRE, Clip skip: 2, ENSD: 31337,
二次元:AnythingV5_v5PrtRE.safetensors [7f96a1a9ca],
2.5D: dreamshaper_8.safetensors [879db523c3],
真实: realisticVisionV60B1_v51VAE.safetensors [15012c538f],
女性写实 xxmix9realistic_v40.safetensors [18ed2b6c48]",

这里我是用了五个基础大模型对其进行测,无论基础模型泛用程度如何,在girl这个提示词下,权重为1的时候都可以做到拟合


人物测试
parameters
sketch,monochrome,greyscale,elderly man,old man,portrait,white_background,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3258169859, Size: 512x512, M
Model: dreamshaper_8.safetensors[879db523c3],
realisticVisionV60B1_v51VAE.safetensors [15012c538f]


经测试该Lora在泛化性较好的模型中表现效果最好,基本可以还原绘画风格和细节,而且在不同权重下的表现具有一定的细节风格差异化,可以通过控制它的使用权重来实现不同风格的展现。


小女孩的年龄特征也能很好识别:
parameters
sketch,monochrome,greyscale,little girl,portrait,white_background,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 4104469418, Size: 512x512, Model hash: 15012c538f,
Model:
realisticVisionV60B1_v51VAE.safetensors [15012c538f]
dreamlikeDiffusion10_10.ckpt [0aecbcfa2c]


道具测试:
parameters
a drawing of a tea pot with a handle,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3652134664, Size: 512x512, Model hash: 15012c538f, Model: realisticVisionV60B1_v51VAE, ENSD: 31337, Style Selector Enabled: True, Style Selector Randomize: False, Style Selector Style: base, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "STRENGTH,0.4,0.6,0.8,1", Y Type: Checkpoint name, Y Values: "realisticVisionV60B1_v51VAE.safetensors [15012c538f],二次元:dreamlikeDiffusion10_10.ckpt [0aecbcfa2c]", Version: v1.8.0
parameters
a drawing of a chicken and her chicks,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3325728720, Size: 512x512, Model hash: 15012c538f, Model: realisticVisionV60B1_v51VAE, Clip skip: 2, ENSD: 31337, Style Selector Enabled: True, Style Selector Randomize: False, Style Selector Style: base, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "STRENGTH,0.4,0.6,0.8,1", Y Type: Checkpoint name, Y Values: realisticVisionV60B1_v51VAE.safetensors [15012c538f], Version: v1.8.0

a scooter is shown in this drawing,
Seed: 2692888502



植物盆栽
parameters
a drawing of a potted plant with leaves,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 2682821664, Size: 640x512, Model hash: 15012c538f, Model: realisticVisionV60B1_v51VAE, ENSD: 31337, Style Selector Enabled: True, Style Selector Randomize: False, Style Selector Style: base, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "STRENGTH,0.4,0.6,0.8,1", Y Type: Checkpoint name, Y Values: realisticVisionV60B1_v51VAE.safetensors [15012c538f], Version: v1.8.0

a drawing of a turtle with a long neck

a drawing of a handbag with a teddy bear in it

a drawing of a cat sitting on the ground

masterpiece,character sketch,a man in a T-shirt puts one hand on his waist and raises his thumb with the other,best details,complete composition,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
masterpiece,Character sketch,a boy in a t - shirt and pants is standing,best details,complete composition,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
masterpiece,sketch,1boy,portrait,monochrome,greyscale,solo,open mouth,white background,male focus,glasses,round eyewear,best details,complete composition,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
parameters
masterpiece,sketch,a middle aged woman,portrait,monochrome,greyscale, solo,short hair,white background,closed eyes,old,old woman,best details,complete composition,<lora:lbc_lbc_Sketch style-000023:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1491541806, Size: 640x960, Model hash: 15012c538f, Model: realisticVisionV60B1_v51VAE, Clip skip: 2, ENSD: 31337,
parameters
a drawing of a shoe with a shoelacee,white_background,<lora:lbc_lbc_Sketch style:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 2593569393, Size: 512x512, Model hash: 2ac767154a, Model: lbc_Realistic, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "STRENGTH,0.4,0.6,0.8,1", Y Type: Checkpoint name, Y Values: "lbc_Realistic.safetensors [2ac767154a],lbc_Simple_v1.0.safetensors [d3725da87c]", Version: v1.7.0

parameters
Character sketch,a drawing of a boy reaching for a tennis ball,<lora:lbc_lbc_Sketch style:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1354253095, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "STRENGTH,0.4,0.6,0.8,1", Y Type: Checkpoint name, Y Values: "lbc_Realistic.safetensors [2ac767154a],lbc_Simple_v1.0.safetensors [d3725da87c]", Version: v1.7.0
parameters
scene sketch, a drawing of a kitchen with a stove and sink,<lora:lbc_lbc_Sketch style:STRENGTH>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1055419347, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "STRENGTH,0.4,0.6,0.8,1", Y Type: Checkpoint name, Y Values: "lbc_Realistic.safetensors [2ac767154a],lbc_Simple_v1.0.safetensors [d3725da87c]", Version: v1.7.0


使用场景测试效果图:
parameters
masterpiece,scene sketch,a man sitting at a table with a pan of food,complete composition,<lora:lbc_lbc_Sketch style:1>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 2981672932, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Lora hashes: "lbc_lbc_Sketch style: 01bcc0cc4859", Version: v1.7.0


下列基于各个年龄阶层的提示词进行稳定的和准确性测试,具体参数如下:
parameters
masterpiece,sketch, a little boy, portrait, monochrome, greyscale, solo, white background, male focus,<lora:lbc_lbc_Sketch style:1>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 3492475913, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Lora hashes: "lbc_lbc_Sketch style: 01bcc0cc4859", Version: v1.7.0
parameters
masterpiece,sketch,a middle aged woman,portrait,monochrome,greyscale, solo,short hair,white background,closed eyes,old,old woman,<lora:lbc_lbc_Sketch style:1>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 842303577, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Lora hashes: "lbc_lbc_Sketch style: 01bcc0cc4859", Version: v1.7.0
parameters
masterpiece,sketch, a elderly man, portrait, monochrome, greyscale, solo, white background, male focus,<lora:lbc_lbc_Sketch style:1>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 690849033, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Lora hashes: "lbc_lbc_Sketch style: 01bcc0cc4859", Version: v1.7.0
parameters
masterpiece,sketch,a little girl,portrait,monochrome,greyscale,solo,smile,short hair,shirt,white background,closed eyes,necktie,glasses,collared shirt,<lora:lbc_lbc_Sketch style:1>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 855994561, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Lora hashes: "lbc_lbc_Sketch style: 01bcc0cc4859", Version: v1.7.0
parameters
masterpiece,sketch, a little boy, portrait, monochrome, greyscale, solo, white background, male focus,<lora:lbc_lbc_Sketch style:1>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 558019312, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Lora hashes: "lbc_lbc_Sketch style: 01bcc0cc4859", Version: v1.7.0
parameters
masterpiece,sketch,a young man,crew cut,portrait,monochrome,greyscale,solo,t-shirt,white background,male focus,<lora:lbc_lbc_Sketch style:1>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 1202962949, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Lora hashes: "lbc_lbc_Sketch style: 01bcc0cc4859", Version: v1.7.0
parameters
masterpiece,sketch,a young girl,long ponytail,portrait,monochrome,greyscale,solo,t-shirt,white background,sexy,male focus,<lora:lbc_lbc_Sketch style:1>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 218280560, Size: 512x768, Model hash: 2ac767154a, Model: lbc_Realistic, Lora hashes: "lbc_lbc_Sketch style: 01bcc0cc4859", Version: v1.7.0
parameters
masterpiece,sketch,monochrome,greyscale,((captivating:1.2)),((stunning:1.3)),<lora:lbc_lbc_Sketch style:1>,
Negative prompt: extra fingers,fewer fingers,(low quality, worst quality:1.4),(bad anatomy),(inaccurate limb:1.2),bad composition,inaccurate eyes,extra digit,fewer digits,(extra arms:1.2),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 3140897935, Size: 896x512, Model hash: 2ac767154a, Model: lbc_Realistic, Denoising strength: 0.27, Hires upscale: 2, Hires steps: 10, Hires upscaler: 4x-UltraSharp, Lora hashes: "lbc_lbc_Sketch style: 01bcc0cc4859", Version: v1.7.0
六、总结:

呼~~,总算到总结部分了,不容易,最近又特忙,所以只能晚上去处理数据集和测试,实际在这次训练的时间真不算长,从数据集的处理到最后完成测试,大约花费了将近3天时间把,其中最繁琐的环节还是训练集的打标工作。
说一说这次训练的一些经验,这次训练过程来说我这边总共得到了10多个表现优异的模型,不过为了更契合训练集的风格,我最后筛选了五个出来,这五个模型其实都算是拟合度非常高的模型了,而且在稳定性上都有非常棒的效果,虽然我只正对了000023这个模型进行了全方位的测试,但是从训练过程的loss值变化来看,其他几个(最后保留下来)的模型应该都有这不错的表现,所以我都选择保留了下来,全部放在文档的附件里面,也算是多一个选择。
000023模型我在测试中分别从它的稳定性,准确性以及泛化性中都进行了全方位的测试,其表现结果如上所示,不管是在人物的年龄识别上,还是在不同道具的识别和输出,甚至是动物植物,场景中都表现的非常不错,特别是人像部分,更是完美,而且对于很多基础大模型都具有很高的支持,即使是泛化性较低的一些大模型,在权重开到”1“的时候还是会还原Lora的风格,进行输出。
总的来说,这个模型我是满意的。

七、附件:
https://www.liblib.art/modelinfo/9d2707e21fd74da88219f9a26b2c34c9

模型已发布在liblib独家发布,欢迎大家去斧正!
最后希望版主给个加精,我要去下载我们的镇坛之宝,哈哈


回复

使用道具 举报

40

主题

477

帖子

5315

积分

高级丹圣

Rank: 13Rank: 13Rank: 13Rank: 13

积分
5315
发表于 2024-3-26 20:06:36 | 显示全部楼层
谢谢,毕业论文get
回复 支持 反对

使用道具 举报

33

主题

709

帖子

7884

积分

高级丹圣

【少女之友】

Rank: 13Rank: 13Rank: 13Rank: 13

积分
7884

开心娱乐节日勋章

发表于 2024-3-27 07:38:25 | 显示全部楼层
谢谢,毕业论文get
回复 支持 反对

使用道具 举报

8

主题

73

帖子

9330

积分

高级丹圣

Rank: 13Rank: 13Rank: 13Rank: 13

积分
9330
发表于 2024-3-27 10:54:26 | 显示全部楼层
楼主,你在哩布的lora无法下载。你的数据集可以分享出来卖灵石,有些人想要比如我
回复 支持 反对

使用道具 举报

0

主题

196

帖子

1061

积分

初级丹圣

Rank: 8Rank: 8

积分
1061
发表于 2024-3-27 13:39:46 | 显示全部楼层
感谢分享
回复

使用道具 举报

0

主题

196

帖子

1061

积分

初级丹圣

Rank: 8Rank: 8

积分
1061
发表于 2024-4-1 09:04:13 | 显示全部楼层
回复

使用道具 举报

1

主题

2

帖子

190

积分

高级丹童

Rank: 2

积分
190
 楼主| 发表于 2024-4-2 09:00:37 | 显示全部楼层
大长卿 发表于 2024-3-27 10:54
楼主,你在哩布的lora无法下载。你的数据集可以分享出来卖灵石,有些人想要比如我 ...

贴个链接
回复 支持 反对

使用道具 举报

5

主题

57

帖子

1149

积分

初级丹圣

Rank: 8Rank: 8

积分
1149
发表于 2024-4-25 16:30:25 | 显示全部楼层
求求训练集
回复 支持 反对

使用道具 举报

0

主题

38

帖子

236

积分

初级丹师

Rank: 3Rank: 3

积分
236
发表于 2024-5-9 10:29:40 | 显示全部楼层
此乃真大神,鉴定完毕!
回复 支持 反对

使用道具 举报

1

主题

18

帖子

209

积分

初级丹师

Rank: 3Rank: 3

积分
209
发表于 2024-5-9 17:20:02 | 显示全部楼层
666,真乃大神,毕业论文。。。
回复 支持 反对

使用道具 举报

QQ|Archiver|手机版|deepfacelab中文网 |网站地图

GMT+8, 2024-12-14 01:59 , Processed in 0.129806 second(s), 36 queries .

Powered by Discuz! X3.4

Copyright © 2001-2020, Tencent Cloud.

快速回复 返回顶部 返回列表