Ponytail LECO Experimental版本v0.1 Alpha (ID: 460138)

Warning: As a LECO it is a failure but it can help give the model the final push to high/low ponytail.

Turns out a couple dozen anime models I tried are unable to do high_ponytail and low_ponytail. SO I thought maybe the concept is just undertrained and made a LECO with it. The result was a complete failure.

Started thinking about it and decided I could theoretically mix a low_ponytail(N15G) and a high_ponytail(golkolea) lora into a model and then train the LECO from it. After all I am sure all models actually know those tags even if unable to express them.

It worked... kind of. the resulting LECO has an absurdly large dead-zone from -1.5 to 1.5 in which it basically does nothing and only working from -1.5 to -2 and 1.5 to 2 after which it burns.

I also Learnt a lot about Ponytails and some of it still applies to the LECO:

• SD has a bias for frontal views so ponytails of any type will commonly only be generated from profiles, backshots and three_quarter_view.

• Anime SD models have an absurdly high bias for higher than average ponytails, not quite high ponytails, but not coming from the middle of the back of the head as they should. Like this /instead of —.

Anyway in short, this LECO has a bunch of caveats:

• 80% possibility of working when there's already ponytail present.

• Needs very high weight to overcome the natural bias of the models so between 1.5 and 2.2

• The concepts of high and low ponytail clash with the normal ponytail and are actually trained pretty heavily making the models start burning at 2.2.

• Very narrow usage band.

• Burns -2.2 ok -1.5 no effect 1.5 ok 2.2 Burns

I am not happy with it but it was an interesting experiment. I might do a proper V1 in the future if i have time.

I will add in the extra resources the LORA I mixed into the model I used to train the LECO with. So if you really need a high or low ponytail you can also use those.

For anyone interested, I was able to give NAI the ability to make High and low ponytails by using kohya's lora merging utility(sd_scripts\networks) merging them like this:

py merge_lora.py --sd_model D:\wifediff\stable-diffusion-webui\models\Stable-diffusion\final-pruned.ckpt --save_to model_lora_merge.safetensors --models D:\wifediff\stable-diffusion-webui\models\Lora\Super_highhigh_ponytail_hair-000002.safetensors D:\wifediff\stable-diffusion-webui\models\Lora\low_ponytail-1.0.safetensors --ratios 0.3 0.5  

描述:

训练词语: ponytail