v1.1

Trained on NoobAI-XL (NAI-XL) Epsilon-pred 1.1-Version and works with it, but probably better pony with ZoinkScoob / Furry NoobAI merge

There are no strength requirements in this version.

The results are probably noisier than v1.0.

It uses "character1 hooves on character2 butt" now.

There are also prompts with "congafollows", "congaleads" and "congacompletes", "offscreen leader", "offscreen completer", but those don't seem to work. The prompts with "follows" and "bipedal leaning" (and "leads" and "completes") may work better, or perhaps you should just list the pony names. Order still doesn't affect anything, although underscores did in a different version.

If you get painting style and don't want it (there is only one picture in that style, cropped variously), try reducing the strength or adding "painting" to the negative prompt, and "my little pony, friendship is magic" in the positive prompt.

v1.0 details are mostly applicable.

v1.0

Important

Works well enough with ZoinkScoob / Furry NoobAI merge, but not with the model it was trained on (NoobAI-XL (NAI-XL) V-Pred-1.0-Version), and I think not really with the Epsilon version of NoobAI-XL.

Use strength around 0.35. Less than 0.3 might not make a conga line, 0.4 and higher become noisy.

Intended syntax (order does not really work): conga line, pony, character1 hooves on character2, character2 hooves on character3, character3 hooves on character4, my little pony, friendship is magic

Training resolutions (try the top ones):

• Resolution,Aspect,Count,Percentage

• 1216x832, 1.46=0.67, 9, 43%

• 1600x640, 2.5=0.4, 7, 33%

• 1344x768, 1.75=0.57, 3, 14%

• 1472x704, 2.09=0.5, 1, 5%

• 1152x896, 1.29=0.8, 1, 5%

Description

This is one of the first working LoRAs I managed to make using OneTrainer.

When OneTrainer generated samples using DDIM (Euler wouldn't work because of vpred or something) with cfg 7.0 at 1216x832, step 3591 (epoch 57) worked best.

But ComfyUI with Euler or Euler A need a low strength even with this step 3024 (epoch 48).

An example of a caption it was trained on: "applejack hooves on fluttershy, rarity hooves on twilight sparkle, twilight sparkle hooves on pinkie pie, fluttershy hooves on rarity, side view, open smile, facing right, blue background, conga line, twilight sparkle with eyes closed, pinkie pie with eyes closed, rarity with eyes closed, applejack with eyes closed, fluttershy with eyes closed, twilight sparkle wearing party hat, applejack wearing party hat, rarity wearing party hat, my little pony: friendship is magic, unicorn twilight"

"facing left" and "facing right" don't seem to do anything.

"derpy hooves" might work poorly because of "hooves".

Prefer 3-4 ponies of the same size.

OneTrainer settings

LoRA rank: 16

LoRA alpha: 8.0

DoRA: off

Optimizer: ADAMW_8BIT

Scheduler: CONSTANT

Learning rate: 0.0003

Batch size: 4

Train text Encoder 1, 2: Off

"force_v_prediction": false,

Masked training was off.

21 captioned and 1 uncaptioned images in the concept, 8 masks, 12 repeats, 63 steps per epoch

At least a few epochs used brightness and contrast augmentation.

License

See the base model's license for what you cannot do even if I allow it: NoobAI License