Update README.md
Browse files
README.md
CHANGED
@@ -2,59 +2,59 @@
|
|
2 |
Model is also on CivitAI: https://civitai.com/models/87254?modelVersionId=258514
|
3 |
## Model Details
|
4 |
### Version 1 parameters:
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
|
31 |
### Version 1.1 parameters:
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
#### Version 1.1 Improvements:
|
59 |
-**Better style consistency**: The model generates in a style closer to the Rhythm Heaven series much more consistently.
|
60 |
1.0 generated a bit more of a detailed style though so if that's what you want you should use that one.
|
|
|
2 |
Model is also on CivitAI: https://civitai.com/models/87254?modelVersionId=258514
|
3 |
## Model Details
|
4 |
### Version 1 parameters:
|
5 |
+
steps_per_image: 50
|
6 |
+
total_images: 49
|
7 |
+
total_steps: ~2400
|
8 |
+
training_model: Anything_V3
|
9 |
+
network_dim: 128
|
10 |
+
network_alpha: 128
|
11 |
+
network_train_on: both
|
12 |
+
learning_rate: 1e-4
|
13 |
+
unet_lr: 0
|
14 |
+
text_encoder _lr: 5e-5
|
15 |
+
lr_scheduler: constant
|
16 |
+
lr_scheduler_num_cycles: 1
|
17 |
+
lr_scheduler_power: 1
|
18 |
+
train_batch_size: 6
|
19 |
+
num_epochs: 6
|
20 |
+
mixed_precision: fp16
|
21 |
+
save_precision fp16
|
22 |
+
save_n_epochs_type: save_every_n_epochs
|
23 |
+
save_n_epochs_type_value: 1
|
24 |
+
resolution: 512
|
25 |
+
max_token_length: 225
|
26 |
+
clip_skip: 2
|
27 |
+
additional_argument: --shuffle_caption --xformers
|
28 |
+
training_hardware: Google Colab Free Tier: Nvidia Tesla T4 GPU
|
29 |
+
training_time: ~45 minutes
|
30 |
|
31 |
### Version 1.1 parameters:
|
32 |
+
steps_per_image: 20
|
33 |
+
total_images: 122 (61 unique images, doubled amount by mirroring them)
|
34 |
+
total_steps: 2440
|
35 |
+
training_model: Any_LoRA
|
36 |
+
optimizer: AdamW
|
37 |
+
network_dim: 128
|
38 |
+
network_alpha: 128
|
39 |
+
network_train_on: both
|
40 |
+
learning_rate: 1e-4
|
41 |
+
unet_lr: 1e-4
|
42 |
+
text_encoder _lr: 5e-5
|
43 |
+
lr_scheduler: constant
|
44 |
+
lr_scheduler_num_cycles: 1
|
45 |
+
lr_scheduler_power: 1
|
46 |
+
train_batch_size: 8
|
47 |
+
num_epochs: 6
|
48 |
+
mixed_precision: bf16
|
49 |
+
save_precision bf16
|
50 |
+
save_n_epochs_type: save_every_n_epochs
|
51 |
+
save_n_epochs_type_value: 1
|
52 |
+
resolution: 768
|
53 |
+
max_token_length: 225
|
54 |
+
clip_skip: 2
|
55 |
+
additional_argument: --xformers
|
56 |
+
training_hardware: RTX 3090
|
57 |
+
training_time: ~1.5 hours (I don't remember exactly)
|
58 |
#### Version 1.1 Improvements:
|
59 |
-**Better style consistency**: The model generates in a style closer to the Rhythm Heaven series much more consistently.
|
60 |
1.0 generated a bit more of a detailed style though so if that's what you want you should use that one.
|