use: 0.1-1.
It is only suitable for 2D models, 2.5D, 3D and real models are very poor.There are enough possibilities for all models to get images that don't match the rendered action, and since the hit rate varies with different checkpoints, it may be necessary to generate some more images.
The Checkpoint version inherits the training set from this series of Lora and can be used to get exactly the same results without using Lora by entering Lora-specific descriptors.
v2.0 Replaced the model for training, adjusted the parameters, the picture effect is cleaner and better than v1.0.
v2.1 & v2.2 are tiny iterative version made on the basis of v2.0, with a weak tendency of overfitting.
v3.0 is close to the 2.5D style, but the chances of getting a correctly described pose are not really high enough to try it with the data from the effects.
v1.0 is a rudimentary version and the results may not be too good, but it will work and perhaps you need to generate many images from which you can pick out a small amount of usable content.
A small amount of face data in the original training data can interfere with the imaging results, please add your own portrait class lora or open ADetailer.
score_9,score_8_up,score_7_up,reverse cowgirl position,hetero,sex from behind,nude,1girl,1boy,lbrc,nsfw,sex,completely nude,uncensored,on bed,girl on top,lying,on back,best quality,highres,masterpiece,from above,mouth mask,
Description
FAQ
Comments (3)
Thanks for your model. This model works well at the pose. Surprisingly, I found that the model changed the style of your base model. And I really like the changed style, that is, the style of this LoRA. Could I ask which painters' works are mainly used to train this LoRA? Thanks a lot.
This series of training images utilized the AI magnification feature of R-ESRGAN 4x+ Anime6B to convert a limited number of low-resolution image materials into 2D format. The sources of these images were low-pixel renderings of the posture found online, which were few in number and unclear. Through operations such as scaling and cropping, the quantity of materials was expanded to meet the minimum training requirement of 25 images.
@tova20200101639 Thanks for your reply. I have seen that the LoRA generated bad fingers and toes sometimes. I think the reason is the low resolution of origin image, and R-ESRGAN 4x+ Anime6B cannot enlarge images losslessly.
It seems that I can't find this painting style. Maybe this is the magic of deep learning.
Good luck.