This is a LORA model for arm restraints trained on 30×5 artificial + internet data for z-image.

It supports a total of 5 variations with an accuracy of about 80%. (The model is overfitted and sometimes fails to distinguish restraint types; variations may occur in face, body shape, and style.)

Weight 0.6~0

Trigger word (sentence)

she is wearing a monoglove_armbinder which binds her arms behind her back.

she is confined in metal_yoke which locks her neck and wrists in a thick pipe.

she is confined,secured in dark oak wooden_yoke which locks her neck and wrists in a thick wood .

she is wearing a latex bolero_straitjacket_front_tie which confines her arms crossed tightly over her chest .

she is wearing a latex bolero_straitjacket_back_tie which confine and binds her arms tightly behind her back

For Devs

The LORA training parameters are as follows.

z-image LORA train adapter used

• rank64,

• batch 6 ,

• steps 6600,

• lr 0.0002

For v2:

Training z-image is difficult, so I plan to try again after the base model is released.

When creating synthetic data, I utilized @Joschek's LORA for post-processing some data. Thank you to @Joschek for creating such a great LORA.