artoria pendragon (swimsuit ruler) release

README.md

@@ -16,3 +16,17 @@ Relevant full character prompt:
 amber \(genshin impact\), brown eyes, brown hair, long hair, red hair ribbon, brown short shorts, leather gloves, pouch, thighhighs, white boots, medium breasts, cleavage cutout, shrug \(clothing\)
 ```
 ![amber (genshin impact)](characters/amber/extras/example.png)
+
+
+### Artoria Pendragon (Swimsuit Ruler) (Fate Grand/Order)
+[Download .safetensors](https://huggingface.co/breakcore2/loras/tree/main/characters/artoria%20pendragon%20\(swimsuit%20ruler\))
+
+[Example Images and Prompts]()
+
+Relevant full character prompt:
+```
+amber \(genshin impact\), brown eyes, brown hair, long hair, red hair ribbon, brown short shorts, leather gloves, pouch, thighhighs, white boots, medium breasts, cleavage cutout, shrug \(clothing\)
+```
+![amber (genshin impact)](characters/artoria%20pendragon%20\(swimsuit%20ruler\)/extras/example.png)
+
+

characters/artoria pendragon (swimsuit ruler)/artoria pendragon (swimsuit ruler) flexible.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b1e7bd44a65b9970d8258464523805fa7b6c1d651d208dcfb9817af234dc641
+size 151074607

characters/artoria pendragon (swimsuit ruler)/artoria pendragon (swimsuit ruler) strict.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23317d5a77154b96b83f554703a04b69e59b284c139b1a78ed8843a1df1e0545
+size 151074607

characters/artoria pendragon (swimsuit ruler)/recommended prompts.txt

@@ -0,0 +1,7 @@
+flexible for custom outfits:
+artoria pendragon, ahoge, blonde hair, green eyes, french braid, pony tail
+strict or flexible for original outfit:
+artoria pendragon \(swimsuit ruler\) \(fate\), ahoge, blonde hair, green eyes, french braid, pony tail, bare shoulders, large breasts, cleavage, clothing cut out, wrist cuffs, detached collar, navel cutout, feather boa, white leotard, blue necktie, blue pantyhose
+additional features:
+thigh strap, pouch
+feather boa

characters/artoria pendragon (swimsuit ruler)/training notes.md

@@ -0,0 +1,31 @@
+## Training Parameters
+Dataset consisted of:
+```
+synthetic non outfit images: 12 repeats * 9 images = 108
+main outfit images from real humans: 5 repeats * 113 images = 565
+synthetic rate: 108/565 = 19%
+```
+
+Training Parameters:
+```
+$learning_rate = 5e-5
+$lr_warmup_ratio = 0.10
+$train_batch_size = 4
+$num_epochs = 3
+$save_every_n_epochs=1
+$scheduler="constant_with_warmup"
+$network_dim=128
+```
+
+Epoch 1 is flexible and epoch 2 is strict version of the model. 
+
+## Points of failure:
+The majority of naturally derived sources of data only consisted of the original outfit. This made it difficult to learn the identity of the character seperately from the outfit they wear. Due to the nature of their outfit, it was difficult to put on alternative outfits without losing all sense of modesty. To address this issue, an initial model was trained and used to generate images featuring alternative costumes with modest designs and a small handful were selected to be included in the dataset for a new model. An arbitary synthetic rate of 20% was chosen and it worked well enough that it did not warrant additional testing for this model. 
+
+We chose to keep the full booru name of the character as part of the tags so that it would be easy to use with the booru autocomplete extension but this includes the word swimsuit in it and probably contributes to alternative outfits having a harder time appearing correctly when using the full name.
+
+## Considerations for the future:
+This model could have benefited from being trained at 768 resolution. The thin diamond shape patterns  that are supposed to cover the blue leggings do not appear. It has been observed that higher resolution training allows for finer details to be learned. It would likely help in creating a consistent symbol on the blue tie as well.
+
+## Learnings:
+Character designs that lean more towards the questionable side get a wide variety of images on the fetish scale. Sorting through those images was not fun. Community scores for images are also biased towards specific fetishes and niches and have to be taken into account when trying to automate image selection based on score. Instead of picking the top 50th percentile of images of the character, break it down into several categories and take the top 50th percentile or just completely reject the category in some cases.