<h1>仅作学习目的。</h1><h3>DO NOT POST YOUR NSFW WORKS HERE.</h3><h3>Notes on usage:</h3><ol><li>LoRA weights around 1.0 should work fine when this model is the only one applied on.</li><li>The new sampler UniPC works well for me at a iterating numbers around 30.</li><li>Try different photo types such as face close-up, portrait and full-body, and different angles like from side, from behind and from above. This model does not give results as good as the last one (<a target="_blank" rel="ugc" href="https://civarchive.com/models/27261/cherrycat">CherryCat | Stable Diffusion LORA | Civitai</a>), for some combinations that are not trained enough (such as from behind + full body), use the hires-fix to improve the results (For example, method=latent, scale=1.5, denoise=0.7).</li></ol><h3>What is the purpose of this model?</h3>This model is a test case on the dataset scale and regularization.Training a model to generate portraits from front is simple, just smash a bunch of portraits you can find to the training set will do the job.However, it will struggle when trying to generate full-body or photos shot from different angles, and won't respond to your prompts well, since the features are not well seperated and learned.There're ways to ameliorate the poor condition when generating faces that are too small in the photos:<ol><li>The simplest way is to set the generating resolution higher, it helps a bit, but might lead to other problems such as weird body ratio.</li><li>Using the hires-fix. It's quite slow, but with the appropriate parameters, it works great.</li><li>Train a better model that are capable of generating small but good faces directly.</li></ol>Regularization:I can't give a clear and exact explanation about what the regularization is. But what I've learnt about regularizaion is that, it somehow works as a guidance for your model to draw faces right. The regularization set will provide certain degrees of information about the face to the body (for example, the head-to-body ratio, or where the head should be).The information from regularization set, will be more or less absorbed into the model, no matter how. This model is a really good example to show that, you'll notice that this model tends to generate big breasts, even though my training set is of 100% pure face photos. Same phenomenon could be found if you have tried my previous model (<a target="_blank" rel="ugc" href="https://civarchive.com/models/27261/cherrycat">CherryCat | Stable Diffusion LORA | Civitai</a>), the training set is of pure face photos too, but it tends to generate small breasts.Dataset:This model is fed with much more data than the last one, a total of 500 regularization photos consists of different angles, postures, head-to-body ratio, and 500 training photos of corresponding faces.Both the regularization and training set are tagged with wd14 tagger, there are definitely some wrong tags poisoning the whole model, such as some photos taken from above are not tagged as "from above", and will result in generating something with weird head-to-body ratio. But it's just impossible for me to check 500 photos and tags one by one.Conclusions:Regularization works fine just like the last one, the model responds well to the prompts like hair color, camera angles, face ratio in photo, etc.But the scale of this dataset is too big and unnecessary, the training took much longer time and larger dataset means more wrong tags and unqualified data.TODO:<ol><li>Do a better selection on the dataset, add more photos from different angles.</li><li>Try a different training parameters, the one I've just trained is clearly over-baked.</li></ol>

Files with SHA256: e80276b5b22dbbaa0c37a9506282f3818158933420f73dee74ab56c7f7831ac3