khanon commited on
Commit
f3e0dc4
1 Parent(s): 4871e4f

cleans up random notes

Browse files
Files changed (3) hide show
  1. shizuko/README.md +2 -0
  2. style training.md +0 -26
  3. training to-do.md +0 -0
shizuko/README.md CHANGED
@@ -26,6 +26,8 @@ For her normal outfit: `two side up, wa maid, japanese clothes, pink kimono, apr
26
 
27
  For her swimsuit outfit: `twintails, swimsuit, pink bikini, frilled bikini, frills, hair flower, fake animal ears`
28
 
 
 
29
  Not all tags may be necessary.
30
 
31
  Her summer alt's cat ears tend to leak into other outfits, but can usually be fixed with negative `fake animal ears`.
 
26
 
27
  For her swimsuit outfit: `twintails, swimsuit, pink bikini, frilled bikini, frills, hair flower, fake animal ears`
28
 
29
+ For her lewd expression: `naughty face, seductive smile, :d, blush`
30
+
31
  Not all tags may be necessary.
32
 
33
  Her summer alt's cat ears tend to leak into other outfits, but can usually be fixed with negative `fake animal ears`.
style training.md DELETED
@@ -1,26 +0,0 @@
1
- style training
2
- Mixer/magnet anon here, I've been doing so many LoRAs for the past week /2 weeks that my eyes hurt, and my SSD space as well.
3
-
4
- I found good settings for concepts if you have a 4090/3090/3090Ti (basically 24GB VRAM).
5
- For 150-200 images (I use 20 repeats), or for 800~ images I used 14 repeats average (it was 5 different concept folders):
6
-
7
- $learning_rate = 0.0001/1e-4 (or 0.0012/1.2e-3)
8
- $lr_warmup_ratio = 0.05
9
- $train_batch_size = 12
10
- $num_epochs = 4
11
- $save_every_n_epochs=1
12
- $scheduler="cosine_with_restarts"
13
- $network_dim=176 (or 192)
14
- $text_encoder_lr=1.5e-5
15
- $unet_lr=0.00015
16
-
17
- You have to be careful to get the concept but not the style, so play with the LR and repeats.
18
-
19
- For characters, the settings are similar but: less repeats, netdim at 96-128 and always 1e-4 LR
20
-
21
- Also, for concepts I didn't prune tags, but I added a first token to help the LoRA. I felt the resutls were better (but obviously you have to use more prompts), but, better results with promtps>not that good results without prompts.
22
- (For characters I still prune the tags that define it though, it is better in that case)
23
-
24
- ------------
25
-
26
- Have I missed any new model that's worth to try to merge/mix and such?
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
training to-do.md DELETED
File without changes