I found it difficult to achieve a pose where the body bends into a bridge-like shape, curving like a bow, with base_v1. This LoRA was trained on 67 training images depicting curved bridge poses from a variety of viewpoints.
Applying the LoRA allows for more flexible and pronounced bridge poses. It can also produce poses where the subject remains standing in a relatively straight posture while bending only the upper body backward, similar to a JoJo pose.
Description
Initial.
dim8/alpha8, AdamW8bit, UNet LR 1e-04, constant 600 + cosine 25(25/200), 512² resolution, 67 images.
FAQ
Comments (3)
Nice! Can't wait to see what other kinds of poses you cook up
This pose is a god damn pain to get right, so thanks.
I'm curious, were there any images in the dataset, where the back remained on the ground and only the lower half of the body is lifted?
Yes. Of the 67 training images, a little over a dozen contain very similar poses (top-down_bottom-up, on_back). Once the LoRA is applied, it already tends to produce a pose close to the body bridge, so starting with something like “lying, on back” might be enough to bring it out.