Type | |
Stats | 3,124 |
Reviews | |
Published | May 23, 2023 |
Base Model | |
Training | Steps: 5,540 Epochs: 4 |
Trigger Words | still_life |
Hash | AutoV2 A436EF6A30 |
still life,Oil paintings,Flowers, fruit,静物Lora
重要:Lora名字被我修改了<lora:KK_still_life_v2:1>,
复制tag的时候,记得把Lora名字修改过来.<lora:still_life_v3:1>
Important: I have changed Lora's name to <lora:KK_still_life_v2:1>.
When copying the tag, remember to change Lora's name over. <lora:still_life_v3:1>
____________________________________________________________________
在测试这个Lora的时候,非常惊喜,好多模型都可以得到很好的展现,
个人比较喜欢用写实的模型来生成静物图片,当然漫画风格的生成也不错,
我测试了十几个模型,基本没有出现崩坏的情况,你也可以自行尝试一下不同的模型,
V1版本传送门:https://civitai.com/models/56374/ins-stylestill-lifesimple-background
V2版本:
这个版本我用了不同的图片训练集,生成出来的物品也非常丰富,
之所以没有合并在一起发布,我觉得这个Lora生成的风格和v1还是有很大不同,
使用说明:
触发词:still_life
Sampler: DPM++ 2M Karras or Dpm++SDE
CFG: 5-9, steps: 30-50
分辨率建议:512或者768,需要丰富的细节,建议超分,
添加Lora+still life就能生成漂亮的静物画,甚至不需要画质描述词,就是这么不讲道理.
当然也可以添加一些画质和光影的描述词:(masterpiece),(best quality),8k, detailed, ultra detail, soothing tones, muted colors, nature lights,realistic ,blurry, solo focus,等,可以自由发挥,
也可以添加:scratch, squiggly, old,vintage,等复古旧物的描述,
可以参考我发布的图片上的描述信息.
有问题可以留言.由于训练集有些图片水印没有去干净,偶尔会生成一些带水印的,影响不大,20张里可能会有一张,
有时间再升级版本,把水印去干净.
_____________________________________________________________________________
still life, Oil paintings, Flowers, fruit, still life Lora
When testing this Lora, I was very surprised by how well many models could be represented.
Personally, I prefer to use realistic models to generate still life images, but of course caricature style generation is also good.
I have tested about a dozen models and there is basically no crashing, you can also try different models yourself.
V1 version portal:https://civitai.com/models/56374/ins-stylestill-lifesimple-background
Version V2.
I used a different image training set for this version, and the resulting items are very rich.
The reason why I didn't combine them in one release is that I think the style of Lora generation is still very different from v1.
Instructions for use.
Trigger word: still_life
Sampler: DPM++ 2M Karras or Dpm++SDE
CFG: 5-9, steps: 30-50
Resolution recommended: 512 or 768, rich detail required, superscoring recommended.
Adding Lora+still life will produce beautiful still life paintings without even needing a quality descriptor, it's that unreasonable.
Of course you can also add some descriptors for quality and light: (masterpiece), (best quality), 8k, detailed, ultra detail, soothing tones, muted colours, nature lights, realistic ,blurry. solo focus, etc., feel free to play around with it.
You can also add: scratch, squiggly, old, vintage, and other vintage descriptions.
Please refer to the descriptions on the pictures I have posted.
If you have any questions, please leave a comment. As some of the images in the training set are not watermarked properly, some of them are generated occasionally with a watermark, but it does not matter much, there may be one out of 20.
I will update the version to remove the watermark when I have time.
__________________________________________________________________________
PS:不知道为啥我传的图片不能自动识别图片参数信息,还要手动添加,好麻烦,我把图片参数信息全部添加到正面描述中了,就这样将就看吧.
(复制全部参数然后点小箭头就可以把参数添加进SD webui中了)
复制全部参数然后点小箭头就可以把参数添加进SD webui中了