Update README.md
Browse files
README.md
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
![](https://huggingface.co/Agreene5/Rhythm_Heaven_Style_LoRA/resolve/main/CivitAIExamples2/formodelcard.png "3 example images")
|
2 |
-
# Rhythm Heaven Style LoRA for Stable Diffusion 1.5
|
3 |
Model is also on CivitAI: https://civitai.com/models/87254?modelVersionId=258514
|
4 |
## Model Details
|
5 |
### Version 1 parameters:
|
@@ -62,6 +62,36 @@
|
|
62 |
**Removed "rhythm_heaven" trigger**: Seems like a style trigger isn't really necessary, removing it just saves a bit of token length.
|
63 |
**Less unprompted black and white generations**: This one isn't as big but I manually added color to some of the training images to get more variety
|
64 |
which consequently means you'll get less black and white generations.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
## Model Description
|
66 |
Trained on humanoid characters from the Rhythm Heaven series (and some from Wario Ware) using AnyLoRA.
|
67 |
Captions were done manually using booru tags.
|
|
|
1 |
![](https://huggingface.co/Agreene5/Rhythm_Heaven_Style_LoRA/resolve/main/CivitAIExamples2/formodelcard.png "3 example images")
|
2 |
+
# Rhythm Heaven Style LoRA for Stable Diffusion 1.5 + SDXL
|
3 |
Model is also on CivitAI: https://civitai.com/models/87254?modelVersionId=258514
|
4 |
## Model Details
|
5 |
### Version 1 parameters:
|
|
|
62 |
**Removed "rhythm_heaven" trigger**: Seems like a style trigger isn't really necessary, removing it just saves a bit of token length.
|
63 |
**Less unprompted black and white generations**: This one isn't as big but I manually added color to some of the training images to get more variety
|
64 |
which consequently means you'll get less black and white generations.
|
65 |
+
### Version 1 (SDXL) parameters:
|
66 |
+
steps_per_image: 20
|
67 |
+
total_images: 122 (61 unique images, doubled amount by mirroring them)
|
68 |
+
total_steps: 7320
|
69 |
+
training_model: anima_pencil-XL
|
70 |
+
optimizer: Adafactor
|
71 |
+
network_dim: 128
|
72 |
+
network_alpha: 1
|
73 |
+
network_train_on: both
|
74 |
+
learning_rate: 1.2e-3
|
75 |
+
unet_lr: 1.2e-3
|
76 |
+
text_encoder _lr: 1.2e-3
|
77 |
+
lr_scheduler: constant
|
78 |
+
lr_scheduler_num_cycles: 1
|
79 |
+
lr_scheduler_power: 1
|
80 |
+
train_batch_size: 5
|
81 |
+
num_epochs: 15
|
82 |
+
mixed_precision: bf16
|
83 |
+
save_precision bf16
|
84 |
+
save_n_epochs_type: save_every_n_epochs
|
85 |
+
save_n_epochs_type_value: 1
|
86 |
+
resolution: 1024
|
87 |
+
max_token_length: 75
|
88 |
+
clip_skip: 2
|
89 |
+
additional_argument: --xformers
|
90 |
+
training_hardware: RTX 3090
|
91 |
+
training_time: ~6 hours
|
92 |
+
#### Version 1 (SDXL) Improvements:
|
93 |
+
**Cleaner looking images**: All of the images used to train this model were upscaled 2x so outputs are less grainy.
|
94 |
+
**Better prompt understanding**: SDXL has a better understanding of prompts so training a LoRA using it as a base makes the LoRA get a better understanding too.
|
95 |
## Model Description
|
96 |
Trained on humanoid characters from the Rhythm Heaven series (and some from Wario Ware) using AnyLoRA.
|
97 |
Captions were done manually using booru tags.
|