alea31415 commited on
Commit
d088e80
1 Parent(s): 5c19cd9

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +43 -2
README.md CHANGED
@@ -7,7 +7,7 @@ tags:
7
  pipeline_tag: text-to-image
8
  ---
9
 
10
- **This model intends to be the ultimate uma-musume model where I try to train as many things about uma-musume as possible into it.**
11
  Ok it's not totally true. Only characters and outfits are trained, ant not locations/objects/running-style(?) or whatever.
12
 
13
 
@@ -16,6 +16,25 @@ Ok it's not totally true. Only characters and outfits are trained, ant not locat
16
  Comming soon... (Still training)
17
 
18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
19
  ## Concepts
20
 
21
  #### List of characters
@@ -154,4 +173,26 @@ Around 60K images containing
154
  - 21377 fan arts
155
  - 270 support cards
156
  - 7822 resused images: From the above three 4019 images are selected to form a few-shot classifier training set, areas centered at face are cropped out which give 3803 images
157
- - 19279 regularization images
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7
  pipeline_tag: text-to-image
8
  ---
9
 
10
+ **This model intends to be the ultimate uma-musume model where I try to train as many things about uma-musume as possible into it.**
11
  Ok it's not totally true. Only characters and outfits are trained, ant not locations/objects/running-style(?) or whatever.
12
 
13
 
 
16
  Comming soon... (Still training)
17
 
18
 
19
+ ## Usage
20
+
21
+ You don't need to use the model as is. Treat it as lora and do either
22
+
23
+ - **Your favorite model + w * (umamusume model - ACertainty/Nai)**
24
+ - or **Umamusume model + w * (your favorite model - ACertainty/Nai)**
25
+
26
+ #### Embeddings
27
+
28
+ For character specific outfit you'd better download embeddings from (coming soon...)
29
+
30
+ #### How to prompt
31
+
32
+ Here are two example captions that are used for training
33
+
34
+ > AgnesDigital; CurrenChan, tracen school uniform; fanart; 2girls, horse girl, multiple girls, character doll, phone, beads, bangs, bow, blush, animal ears, school uniform, holding phone, two side up, heart in mouth, one eye closed, white background, sweat, tracen school uniform, selfie, red bow, shirt, sailor collar, open mouth, simple background, heart, horse ears, holding, purple shirt, purple skirt, skirt, long sleeves
35
+
36
+ > CurrenChan; SmartFalcon, SmartFalcon-racing-suit; fanart; 2girls, horse girl, multiple girls, dress, diamond (shape), party, nail polish, red bow, see-through, teeth, frilled dress, upper teeth only, balloon, blurry, frilled collar, animal ears, timestamp, ring, black bow, pink dress, border, gem, short sleeves, puffy sleeves, heart hands, pink nails, strap, bow, lace-trimmed sleeves, black headband, chromatic aberration, white border, black dress, writing on wall, headband, frills, horse tail, chain, horse ears, puffy short sleeves, bracelet, tail, bangs, open mouth, collar, black nails, multicolored dress, multicolored clothes, depth of field, heart
37
+
38
  ## Concepts
39
 
40
  #### List of characters
 
173
  - 21377 fan arts
174
  - 270 support cards
175
  - 7822 resused images: From the above three 4019 images are selected to form a few-shot classifier training set, areas centered at face are cropped out which give 3803 images
176
+ - 19279 regularization images
177
+
178
+
179
+ ## Training
180
+
181
+ The model is trained in two phases (resolution 512, clip skip 1) on top of [ACertainty](https://huggingface.co/JosephusCheung/ACertainty)
182
+
183
+ **The first phase is trained with [EveryDream2](https://github.com/victorchall/EveryDream2trainer)**
184
+
185
+ - Specific weighting scheme with maximum repeat 50
186
+ - consine scheduler, 2.5e-6 learning rate
187
+ - conditional dropout 0.08
188
+
189
+ Total steps x batch is around 1.1M
190
+ - batch 8 for 61495 steps
191
+ - batch 6 for ~3500 steps but this gets stopped due to some technical issues
192
+ - batch 4 for 150621 steps
193
+
194
+ **The second phase is trained with [naifu trainer](https://github.com/AdjointOperator/naifu-diffusion)**
195
+
196
+ The goal is to train embeddings for character-specific outfits.
197
+ Both unets and embeddings are trained.
198
+ This phase does not use the anime screenshots nor the regularization images.