Pretzel Start

2025.09.28 Update - added WAN2.2 i2v version

Hi, this is Image To Video lora for pretzel pose (legs behind head).

This is very hard pose so you need to try different seeds with different combination of parameters to get acceptable results and select a start image where a model is sitting/lying on something large enough and stable, is turned towards the viewer and her hands are pressed to her body. I hope to improve it in Wan 2.2.

Trained on 19 videos 1-6 seconds length.

Trained resolution 832x464, 171 epochs.

For generation I’m using Wan14BT2VFusionX_fp16 model for better results with the following parameters:

Steps: 10-12

Sampler: euler simple, euler ancestral beta

Shift: 4-8

Resolution I tried up to 1280x720 with good results.

Trigger word: pretzel_pose

Sample phrases:

pretzel_pose, a nude woman is lying on a mat, she starts doing pretzel pose, she puts her legs behind her head, in a garden, front view

pretzel_pose, a nude woman is sitting on a chair, she is wearing black high heels, she starts doing pretzel pose, she puts her legs behind her head, in a photo studio, front view, static camera

pretzel_pose, a nude woman is lying on a bed, she is wearing sheer stockings. she starts doing pretzel pose, she puts her legs behind her head, in a bedroom, front view, static camera

Check out attached videos for detailed prompts.

P.S. As I understood the I2V model is more practical than T2V since you can generate a start image more precisely using any preferable model, so I2V models training will have more priority, but we can discuss it in comments.