Read the description first

This is the lora version of thirst trap girl, still free to download.

This is a short video style lora model, after the short video platform intelligent algorithms screened out more suitable for the national aesthetic beauty and handsome, while the model out of the figure is also more suitable for use in the Tusheng video, more in line with the aesthetic of the short video, you say is it.

This version focuses on enhancing the learning of face, body type, skin color, and jitterbug style.

You can either fill in the trigger word or dygirl.

This model has screened 500 popular short videos in the past two years (varies from person to person), and the last remaining 100 videos for frame extraction. Most of them use 3-5 frames/second for extraction, eliminating the vignette and the interference part of the picture, and then repairing the hand of each picture (dynamic frame hand is almost always blurred) to arrive at the training set (in order to ensure the training effect, the same person's picture are censored).

This version basically solves the problem of hand degradation in the distillation model, but the training set is obtained by extracting frames and repairing hands, so it is inevitable that there are some defects.

The example graphs are all non-training set of graph backpropagation tags, cue words in the node “show text”. Click on the example graph, check the workflow, and find the show text node.

This version recommends using comfyui for best results.