khanon commited on
Commit
b8e040d
1 Parent(s): eef5939

adds retrained Sora LoRA

Browse files
README.md CHANGED
@@ -6,16 +6,54 @@ license: mit
6
 
7
  Here you will find the various LoRAs I've trained, typically of Blue Archive characters.
8
 
9
- ## 🚧 IMPORTANT 🚧
10
-
11
- I haven't moved all of my LoRAs over from Mega. Please see the [Mega.co.nz repository](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA) if a LoRA is missing from here.
12
-
13
  ## Blue Archive
14
  ブルーアーカイブ / 블루 아카이브
15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16
  ### Shizuko
17
  Kawawa Shizuko / 河和シズコ / 카와와 시즈코
18
 
19
  [![Shizuko](shizuko/chara-shizuko.png)](https://huggingface.co/khanon/lora-training/blob/main/shizuko/README.md)
20
 
 
 
21
 
 
 
6
 
7
  Here you will find the various LoRAs I've trained, typically of Blue Archive characters.
8
 
 
 
 
 
9
  ## Blue Archive
10
  ブルーアーカイブ / 블루 아카이브
11
 
12
+ ### Atsuko
13
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
14
+
15
+ ### Chise
16
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
17
+
18
+ ### Hibiki
19
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
20
+
21
+ ### Hina
22
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
23
+
24
+ ### Iroha
25
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
26
+
27
+ ### Izuna
28
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
29
+
30
+ ### Kokona
31
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
32
+
33
+ ### Mari
34
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
35
+
36
+ ### Michiru
37
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
38
+
39
+ ### Miyako (WIP)
40
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
41
+
42
+ ### Mutsuki
43
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
44
+
45
+ ### Natsu
46
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
47
+
48
+ ### Resia
49
+ [Available on old Mega.co.nz repository.](https://mega.nz/folder/SqYwQTRI#GN2SmGTBsV6S4q-L-V4VeA)
50
+
51
  ### Shizuko
52
  Kawawa Shizuko / 河和シズコ / 카와와 시즈코
53
 
54
  [![Shizuko](shizuko/chara-shizuko.png)](https://huggingface.co/khanon/lora-training/blob/main/shizuko/README.md)
55
 
56
+ ### Sora
57
+ Sora / ソラ (ブルーアーカイブ) / 소라 (블루 아카이브)
58
 
59
+ [![Sora](sora/chara-sora-v3.png)](https://huggingface.co/khanon/lora-training/blob/main/sora/README.md)
sora/README.md CHANGED
@@ -1,30 +1,51 @@
1
- # Sora (Blue Archive NPC)
2
- Decided to try completely changing things up after seeing an anon's suggestions on /hdg/. Used batch size 12, lower resolution, and a high learning rate. Can't tell if it's any better or worse quality-wise than my usual LoRAs, but it was a lot faster to train.
3
 
4
- Be forewardned that this LoRA is prone to generating NSFW results since Sora an unusually high proportion of very lewd art.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5
 
6
  ## Usage
7
- Didn't do much tag pruning so you have to prompt a lot.
8
 
9
  Use any or all of the following tags to summon Sora: `sora, halo, mini wings, white wings, two side up, blonde hair, forehead`
10
 
11
- For her usual outfit, `blue apron, white shirt` is typically enough. You can also add `off-shoulder, strap slip, bowtie, short sleeves` if desired.
 
 
12
 
13
- Sometimes her wings appear too high; `low wings` can help with this. Her wings are also usually a bit too big but that's kind of hard to control.
14
 
15
- Weight 1 seems to work fine.
16
 
17
  ## Training
18
- *All parameters are provided in the accompanying JSON files.*
19
- - Trained on a set of 128 images, repeated 25 times, 3 epochs (128 images * 25 repeats / 12 batch size * 3 epochs = 800 steps)
20
- - Dataset included a mixture of SFW and NSFW.
21
- - Batch size 12 is up from usual of 3
22
- - Resolution was 512, down from usual of 832
23
- - General learning rate is 1.2e-3, significantly up from the Kohya default (1e-6 I believe)
24
- - My understanding is that when this learning rate is set at the same time as the text encoder and unet rates, it is used only as one of the inputs to the AdamW optimizer, which adjusts the learning rate as necessary. So even though this value is high, it is not running at this value constantly.
25
- - Text encoder learning rate of 1.5e-5
26
- - Unet learning rate of 1.5e-4
27
  - `constant_with_warmup` scheduler instead of `cosine`
28
  - Initially tagged with WD1.4 Convnextv2 model. Tags minimally pruned/edited.
29
- - Used network_dimension 128 (same as usual) and network_alpha 64
30
  - Trained without VAE.
 
 
 
 
 
 
 
 
 
1
+ # Sora (Blue Archive)
2
+ ソラ (ブルーアーカイブ) / 소라 (블루 아카이브)
3
 
4
+ Be forewarned that this LoRA is prone to generating NSFW results since Sora an unusually high proportion of extremely lewd art. Use `nsfw, naked apron` etc in the negative prompt to prevent this.
5
+
6
+ [**Download here.**](https://huggingface.co/khanon/lora-training/blob/main/sora/chara-sora-v3.safetensors)
7
+
8
+ ## Table of Contents
9
+ - [Preview](#preview)
10
+ - [Usage](#usage)
11
+ - [Training](#training)
12
+ - [Revisions](#revisions)
13
+
14
+ ## Preview
15
+
16
+ ![Sora portrait](chara-sora-v3.png)
17
+ ![Sora](example-001-v3-defmix1.png)
18
+ ![Sora fallen](example-002-v3-aom2nmm.png)
19
 
20
  ## Usage
 
21
 
22
  Use any or all of the following tags to summon Sora: `sora, halo, mini wings, white wings, two side up, blonde hair, forehead`
23
 
24
+ For her usual outfit: `blue apron, white shirt, (black pleated skirt:0.8)`
25
+ - If needed, you can also add `off-shoulder, strap slip, bowtie, short sleeves`
26
+ - Sora's skirt is not visible in most training art. Consider adding `black pleated skirt` and negative prompt `miniskirt, pencil skirt` for more accuracy.
27
 
28
+ Sometimes her wings appear too high; `low wings` can help with this.
29
 
30
+ Weight 1 is suggested.
31
 
32
  ## Training
33
+ *Exact parameters are provided in the accompanying JSON files.*
34
+ - Trained on a set of 128 images; 88 swimsuit, 37 normal.
35
+ - Datset included a mixture of SFW/NSFW. Sora has more NSFW than most students.
36
+ - 9 repeats
37
+ - 3 batch size, 4 epochs
38
+ - `(128*9) / 3 * 4` = 1536 steps
39
+ - 768x768 training resolution
 
 
40
  - `constant_with_warmup` scheduler instead of `cosine`
41
  - Initially tagged with WD1.4 Convnextv2 model. Tags minimally pruned/edited.
42
+ - Used network_dimension 128 (same as usual) / network alpha 128 (default)
43
  - Trained without VAE.
44
+
45
+ ## Revisions
46
+ - v3 (2023-02-06)
47
+ - Completely re-trained to correct issue where caption files were not being loaded correctly. New version should have more flexibility but requires additional outfit prompting.
48
+ - v2b (2023-02-01)
49
+ - Initial release.
50
+ - **Issue:** caption files were completely ignored during training. As a result, prompting "sora" grants a very strong effect, but the LoRA is highly overfit.
51
+ - [Download available on Mega.co.nz repository.](https://mega.nz/folder/yjwGiIBD#w79CxW2-Cj66QmObg0zOdg)
sora/chara-sora-v3.png ADDED

Git LFS Details

  • SHA256: 238617e2237c49f319ec12362bb22c99be35290b5a65726afa0fed478cda172a
  • Pointer size: 131 Bytes
  • Size of remote file: 429 kB
sora/chara-sora-v3.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:14f8128b46cba1bea83f56316434ea70dc5d3fd17547d6328ff4b30c6309f4b4
3
+ size 151120215
sora/example-001-v3-defmix1.png ADDED

Git LFS Details

  • SHA256: 13da455afa66d687f952c6688a42e25d6d5e0953bdce6427e350680eb2ea9422
  • Pointer size: 132 Bytes
  • Size of remote file: 2.17 MB
sora/example-002-v3-aom2nmm.png ADDED

Git LFS Details

  • SHA256: af3f4850e32d7008e1c4aed04f3204e54274574f08b54c85bf3242078fc19c07
  • Pointer size: 132 Bytes
  • Size of remote file: 2.22 MB
sora/lora_chara_sora_v3_128i9r.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "pretrained_model_name_or_path": "G:/sd/repo/models/Stable-diffusion/nai-animefull-final-pruned.safetensors",
3
+ "v2": false,
4
+ "v_parameterization": false,
5
+ "logging_dir": "",
6
+ "train_data_dir": "G:/sd/training/datasets/sora/dataset",
7
+ "reg_data_dir": "",
8
+ "output_dir": "G:/sd/lora/trained/chara/sora",
9
+ "max_resolution": "768,768",
10
+ "learning_rate": "1e-5",
11
+ "lr_scheduler": "constant_with_warmup",
12
+ "lr_warmup": "5",
13
+ "train_batch_size": 3,
14
+ "epoch": "4",
15
+ "save_every_n_epochs": "",
16
+ "mixed_precision": "fp16",
17
+ "save_precision": "fp16",
18
+ "seed": "31337",
19
+ "num_cpu_threads_per_process": 32,
20
+ "cache_latents": true,
21
+ "caption_extension": ".txt",
22
+ "enable_bucket": true,
23
+ "gradient_checkpointing": false,
24
+ "full_fp16": false,
25
+ "no_token_padding": false,
26
+ "stop_text_encoder_training": 0,
27
+ "use_8bit_adam": true,
28
+ "xformers": true,
29
+ "save_model_as": "safetensors",
30
+ "shuffle_caption": true,
31
+ "save_state": false,
32
+ "resume": "",
33
+ "prior_loss_weight": 1.0,
34
+ "text_encoder_lr": "1.5e-5",
35
+ "unet_lr": "1.5e-4",
36
+ "network_dim": 128,
37
+ "lora_network_weights": "",
38
+ "color_aug": false,
39
+ "flip_aug": false,
40
+ "clip_skip": 2,
41
+ "gradient_accumulation_steps": 1.0,
42
+ "mem_eff_attn": false,
43
+ "output_name": "chara-sora-v3",
44
+ "model_list": "",
45
+ "max_token_length": "150",
46
+ "max_train_epochs": "",
47
+ "max_data_loader_n_workers": "",
48
+ "network_alpha": 128,
49
+ "training_comment": "128 images * 9 repeats / 3 batch size * 4 epochs = 1536 steps",
50
+ "keep_tokens": 2,
51
+ "lr_scheduler_num_cycles": "",
52
+ "lr_scheduler_power": "",
53
+ "persistent_data_loader_workers": true
54
+ }