Create v2.0.md
Browse files- hisakawa/release_notes/v2.0.md +100 -0
hisakawa/release_notes/v2.0.md
ADDED
@@ -0,0 +1,100 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# v2.0
|
2 |
+
|
3 |
+
A LoRA of the Hisakawa twins from THE iDOLM@STER Cinderella Girls based on 967 fanarts. More style neutral than v1.1.
|
4 |
+
|
5 |
+
Two versions are available: rank 1 and rank 4.
|
6 |
+
|
7 |
+
This shows that, perhaps unsurprisingly, 1 rank can carry info of more than 1 character.
|
8 |
+
|
9 |
+
Despite the abundance of two-subject data in the training set, the model still struggles to seperate the attributes of the twins in the image. Increasing the model rank may or may not help. This will be investigated some time in the future.
|
10 |
+
|
11 |
+
## Usage
|
12 |
+
|
13 |
+
The character tags are `Hisakawa Nagi` and `Hisakawa Hayate`.
|
14 |
+
|
15 |
+
To generate both characters in the same image, use either `Hisakawa Nagi / Hisakawa Hayate` or `Hisakawa Hayate / Hisakawa Nagi`
|
16 |
+
|
17 |
+
|
18 |
+
Top related tags:
|
19 |
+
```
|
20 |
+
Hiwakawa Nagi,
|
21 |
+
1girl, solo, twintails, long hair, grey hair, braid, low twintails, brown eyes, looking at viewer, ribbon, hair ribbon, braided bangs, blush, shirt, simple background, long sleeves, white background, black ribbon, very long hair, bow, white shirt, jacket, parted lips, skirt, collared shirt, closed mouth, thighhighs, hair over shoulder, :o, upper body, puffy sleeves, shorts, open mouth, breasts, red bow, black jacket, collarbone, short sleeves, holding, small breasts
|
22 |
+
|
23 |
+
Hisakawa Hayate,
|
24 |
+
1girl, solo, long hair, grey hair, braid, looking at viewer, blush, smile, jewelry, breasts, braided bangs, earrings, blue eyes, very long hair, white background, shirt, open mouth, simple background, skirt, collarbone, medium breasts, white shirt, long sleeves, cleavage, pleated skirt, jacket, collared shirt, :d, bow, bare shoulders, navel, swimsuit, closed mouth, hair between eyes, holding, upper body, puffy sleeves, dress, bikini, short sleeves
|
25 |
+
|
26 |
+
Hisakawa Nagi / Hisakawa Hayate,
|
27 |
+
2girls, multiple girls, siblings, twins, braid, long hair, sisters, grey hair, twintails, brown eyes, smile, braided bangs, open mouth, blue eyes, low twintails, ribbon, blush, looking at viewer, hair ribbon, jewelry, earrings, shirt, dress, skirt, long sleeves, black ribbon, very long hair, white shirt, bow, parted lips, breasts, :d, simple background, one eye closed, v, collarbone, white background, :o, puffy sleeves, collared shirt
|
28 |
+
```
|
29 |
+
|
30 |
+
For specific outfits, refer to the preview images or the dataset.
|
31 |
+
|
32 |
+
## Training info
|
33 |
+
|
34 |
+
Dataset: [967 fanarts from Danbooru](https://huggingface.co/datasets/gustproof/sd-data/blob/main/hisakawa.zip)
|
35 |
+
* 338 Hisakawa Nagi solo
|
36 |
+
* 341 Hisakawa Hayate solo
|
37 |
+
* 238 both twins
|
38 |
+
|
39 |
+
No balancing and tag cleaning.
|
40 |
+
|
41 |
+
For images with both characters, the order of the character tags were shuffled dynamically.
|
42 |
+
|
43 |
+
|
44 |
+
Training cost: ~2 T4-hour each for r1 and r4
|
45 |
+
|
46 |
+
|
47 |
+
Training config:
|
48 |
+
```
|
49 |
+
[model_arguments]
|
50 |
+
v2 = false
|
51 |
+
v_parameterization = false
|
52 |
+
pretrained_model_name_or_path = "animefull-final-pruned.ckpt"
|
53 |
+
|
54 |
+
[additional_network_arguments]
|
55 |
+
no_metadata = false
|
56 |
+
unet_lr = 0.0001
|
57 |
+
text_encoder_lr = 0.0001
|
58 |
+
network_module = "networks.lora"
|
59 |
+
network_alpha = 1
|
60 |
+
network_train_unet_only = false
|
61 |
+
network_train_text_encoder_only = false
|
62 |
+
|
63 |
+
[optimizer_arguments]
|
64 |
+
min_snr_gamma = 5
|
65 |
+
optimizer_type = "AdamW8bit"
|
66 |
+
learning_rate = 0.0001
|
67 |
+
max_grad_norm = 1.0
|
68 |
+
lr_scheduler = "constant"
|
69 |
+
lr_warmup_steps = 0
|
70 |
+
|
71 |
+
[dataset_arguments]
|
72 |
+
debug_dataset = false
|
73 |
+
dataset_repeats = 1
|
74 |
+
shuffle_caption = true
|
75 |
+
keep_tokens = 1
|
76 |
+
resolution = "512,512"
|
77 |
+
caption_dropout_rate = 0
|
78 |
+
caption_tag_dropout_rate = 0
|
79 |
+
caption_dropout_every_n_epochs = 0
|
80 |
+
color_aug = false
|
81 |
+
token_warmup_min = 1
|
82 |
+
token_warmup_step = 0
|
83 |
+
|
84 |
+
[training_arguments]
|
85 |
+
save_precision = "fp16"
|
86 |
+
save_every_n_epochs = 1
|
87 |
+
train_batch_size = 4
|
88 |
+
max_token_length = 225
|
89 |
+
mem_eff_attn = false
|
90 |
+
xformers = true
|
91 |
+
max_train_epochs = 9999
|
92 |
+
max_data_loader_n_workers = 8
|
93 |
+
persistent_data_loader_workers = true
|
94 |
+
gradient_checkpointing = false
|
95 |
+
gradient_accumulation_steps = 1
|
96 |
+
mixed_precision = "fp16"
|
97 |
+
clip_skip = 2
|
98 |
+
lowram = true
|
99 |
+
|
100 |
+
```
|