The pose is
• Knees and toes turned inwards, AKA a knock-knee and pigeon-toe.
• This changes leg pose, not arm pose
• Works for:
• standing or sitting
• with shoes or barefoot
• front, side, or back view
• all ethnicities and body sizes
• females only
Krea2 v1
Usage:
• Strength: 1.0 works with no distortion most of the time. If you get distortion, use lower strength.
• Prompting:
• No trigger or special prompt is needed for the legs
• Tell the arms what to do, this lora doesn't impact them
• "feet apart" - gives a more X shaped leg pose
• "toes together" and/or "knees together" - for stubborn cases
• "shy" or "shy pose" - not trained, but natively changes overall posture, arm and hand position
Limitations:
• Mostly doesn't work with males, but will at very high strength
Training method:
• 39 images, hand edited captions, all full body with all faces obscured (zero faces in the dataset), ~50/50 standing and sitting
• 50% synthetic images from earlier versions. This was essential to get variation that I couldn't find, especially barefoot images
• ~4k steps, 0.0003 static LR, 32 rank. Then I pruned unneeded blocks and boosted some block strength with Fizgig.
ZiT version (not recommended)
Usage:
• Start your prompt with the full phrase, "A woman is standing in a shy pose with her toes pointing together."
• Use strength 1.3.
Limitations, sadly:
• Only one person
• No males
• No side profile
• No Sitting or lying down poses
• Impacts faces and other aspects of the image
Training method:
• Trained for $2 on a service that had no settings
• I used 20 images: legs and feet only (I learned later this is a bad technique)