Update (v0.4)

The new version was trained longer with more images of higher quality and with a larger training resolution (increased from 1024 to 1280). It has far fewer anatomical issues than v0.3.

This is a style lora trained on screencaps from MMD videos

The trigger word is 'mmdstyle2'

Recommended weight: 1.0

Older Versions (v0.3)

The trigger word is 'mmdstyle'

Recommended weight: ~0.7

Known Issues (v0.3)

Sometimes hands or arms are missing if the pose is not specified.

This can be mitigated by scheduling the lora or trigger word to trigger after a few initial steps.

A1111 syntax example: [(mmdstyle:0.1):(mmdstyle:1.0):0.5]

(This sets the trigger word 'mmdstyle' to 10% strength for the first 50% of your steps, then increases it to 100% strength in the last half, causing it to influence style but not composition)