Trained on 16 images, mostly close-up portrait of Chinese actress/pop singer 杨超越 Yang Chaoyue. Processed with focal point crop. Weight this embedding at 0.8~1.2 is suggested. See sample images and png info to grab some prompts that just works.
Hires.fix on is suggested (e.g. 512*768--2x-->1024*1536, Hires steps 18,Denoising strength 0.15, ESRGAN_4x or LDSR are highly suggested as always, depending on trade-off between speed and super-sharp details.)
Try out Restore Faces when generating side view photos since the training set isn't varied enough concerning poses and angle, though it performs quite well on contrapposto or selfie photos.
中文翻译 Simplified Chinese Translation:
训练了 16 张图像,大部分是中国女演员/流行歌手 杨超越的特写肖像。使用焦点裁剪处理。建议将此嵌入权重设置为 0.8~1.2。查看示例图像和 png 信息以获取一些有效的提示。
建议启用 Hires.fix超分辨率(例如 512*768--2x-->1024*1536,Hires 步骤 18,去噪强度 0.15,一如既往地强烈建议 ESRGAN_4x 或 LDSR,取决于你在速度和细节之间的权衡)
在生成侧视图照片时尝试 Restore Faces,因为训练集在姿势和角度方面的变化不够大,尽管它在 contrapposto 或自拍照片上表现相当不错。
请勿用于商业用途或生成违法/侵权照片用于传播。
Description
Initial commit, trained at 4000 steps, batch size 16.
FAQ
Comments (6)
woshele
练度有点低吧,感觉还不太像。
Wgich models do you recommend with this TI ?
你好,请问你是用webui训练的embedding的么。我这边训练了一个程潇的Lora,你可以在我的主页查看,但是embedding一直失败,不知道为何。
咋用啊扔到stable-diffusion里用不了
怎么感觉有点像Jackey love
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.