v70 for fluffyrock and its mixed models (especially for indigofurrymix)
v66 for anime models with sufficient NAI components
The trigger words I gave are NECESSARY
v66 sample pictures above are from a mixed model, 【indigo16 0.5counterfeit】 0.5【furryvixen2 0.3counterfeit】 . You can get it at my 🤗https://huggingface.co/FrixH2022/counindivixen
/////////////////////
v70适配fluffyrock和他的一些混合模型（特别是indigofurrymix，专门给indigo做适配的）
v66适配NAI成分比较多的二次元模型
我发的触发词比较重要，因为我选择了不删标训练，它们是在训练集中出现频率很多的词。当然如果你想扣掉某个人物特征也可以自己修改触发词
v66的示例图的底模是先indigo16和0.5的counterfeit3混、furryvixen2和0.3的counterfeit3，再两个0.5混起来。你也可以直接在这里下https://huggingface.co/FrixH2022/counindivixen
超喜欢士郎~
Description
180 pics trained on NAI
FAQ
Comments (4)
Viewed your twitter, I suggest leaving some pure generated images for people to know the prompts used. Very good LoRAs though, hope someday you could share how did you train them.
I used the 3 default learning rates and 8bit optimizer to train these character loras. The datasets require high-quality images with minimize interference elements and the image scale can be arbitrary, but the resolution is better not less than 768*768 (also 768*768 is selected in the training settings). And you can use Tencent ARC to enlarge the small images. keep token 1 (needs one trigger word),
batch size 6 on RTX3090 24GB，always get satisfactory results at 12 epochs
Other parameters are almost default
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.