Bernard Maltais
commited on
Commit
•
3f2a7cc
1
Parent(s):
716a8e7
1st commit
Browse files- 1.png +0 -0
- 2.png +0 -0
- 3.png +0 -0
- README.md +28 -0
- by_laze_opera_panda.ckpt +3 -0
- kohya_diffusers_train.ps1 +80 -0
1.png
ADDED
![]() |
2.png
ADDED
![]() |
3.png
ADDED
![]() |
README.md
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: unknown
|
3 |
+
---
|
4 |
+
# Soda Stream finetuned style Model
|
5 |
+
|
6 |
+
Produced from publicly available pictures in landscape, portrait and square format.
|
7 |
+
|
8 |
+
## Model info
|
9 |
+
|
10 |
+
The models included was trained on "multi-resolution" images.
|
11 |
+
|
12 |
+
## Using the model
|
13 |
+
|
14 |
+
* common subject prompt tokens: `<wathever> by laze opera panda`
|
15 |
+
|
16 |
+
## Example prompts
|
17 |
+
|
18 |
+
`woman near a fountain by laze opera panda`:
|
19 |
+
|
20 |
+
<img src="https://huggingface.co/cyburn/laze_opera_panda/resolve/main/1.png" alt="Picture." width="500"/>
|
21 |
+
|
22 |
+
`woman in taxi by laze opera panda`:
|
23 |
+
|
24 |
+
<img src="https://huggingface.co/cyburn/laze_opera_panda/resolve/main/2.png" alt="Picture." width="500"/>
|
25 |
+
|
26 |
+
`man portrait by laze opera panda`:
|
27 |
+
|
28 |
+
<img src="https://huggingface.co/cyburn/laze_opera_panda/resolve/main/3.png" alt="Picture." width="500"/>
|
by_laze_opera_panda.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9901f0f9fdcf909bcb502d95f7c5e9c69a05457bc50c7f88efa105ee6fd44eaf
|
3 |
+
size 2132856686
|
kohya_diffusers_train.ps1
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# by soda stream
|
2 |
+
|
3 |
+
# variable values
|
4 |
+
$pretrained_model_name_or_path = "D:\models\v1-5-pruned-mse-vae.ckpt"
|
5 |
+
$train_dir = "D:\dreambooth\train_laze_opera_panda"
|
6 |
+
$training_folder = "raw"
|
7 |
+
|
8 |
+
$learning_rate = 1e-6
|
9 |
+
$dataset_repeats = 40
|
10 |
+
$train_batch_size = 8
|
11 |
+
$epoch = 4
|
12 |
+
$save_every_n_epochs=1
|
13 |
+
$mixed_precision="bf16"
|
14 |
+
$num_cpu_threads_per_process=6
|
15 |
+
|
16 |
+
$max_resolution = "576,576"
|
17 |
+
|
18 |
+
# You should not have to change values past this point
|
19 |
+
|
20 |
+
# stop script on error
|
21 |
+
$ErrorActionPreference = "Stop"
|
22 |
+
|
23 |
+
# activate venv
|
24 |
+
.\venv\Scripts\activate
|
25 |
+
|
26 |
+
# Usefull to create base caption that will be augmented on a per image basis
|
27 |
+
|
28 |
+
$caption="by laze opera panda"
|
29 |
+
|
30 |
+
$files = Get-ChildItem $train_dir\$training_folder\"*.*" -Include *.png,*.jpg,*.webp
|
31 |
+
foreach ($file in $files) {New-Item -ItemType file -Path $train_dir\$training_folder -Name "$($file.BaseName).txt" -Value $caption}
|
32 |
+
|
33 |
+
|
34 |
+
# create caption json file
|
35 |
+
python D:\kohya_ss\diffusers_fine_tuning\merge_captions_to_metadata-ber.py `
|
36 |
+
--caption_extention ".txt" $train_dir"\"$training_folder $train_dir"\meta_cap.json"
|
37 |
+
|
38 |
+
# create images buckets
|
39 |
+
python D:\kohya_ss\diffusers_fine_tuning\prepare_buckets_latents-ber.py `
|
40 |
+
$train_dir"\"$training_folder `
|
41 |
+
$train_dir"\meta_cap.json" `
|
42 |
+
$train_dir"\meta_lat.json" `
|
43 |
+
$pretrained_model_name_or_path `
|
44 |
+
--batch_size 4 --max_resolution $max_resolution --mixed_precision fp16
|
45 |
+
|
46 |
+
# Get number of valid images
|
47 |
+
$image_num = Get-ChildItem "$train_dir\$training_folder" -Recurse -File -Include *.npz | Measure-Object | %{$_.Count}
|
48 |
+
$repeats = $image_num * $dataset_repeats
|
49 |
+
|
50 |
+
# calculate max_train_set
|
51 |
+
$max_train_set = [Math]::Ceiling($repeats / $train_batch_size * $epoch)
|
52 |
+
|
53 |
+
accelerate launch --num_cpu_threads_per_process $num_cpu_threads_per_process D:\kohya_ss\diffusers_fine_tuning\fine_tune.py `
|
54 |
+
--pretrained_model_name_or_path=$pretrained_model_name_or_path `
|
55 |
+
--in_json $train_dir"\meta_lat.json" `
|
56 |
+
--train_data_dir=$train_dir"\"$training_folder `
|
57 |
+
--output_dir=$train_dir"\fine_tuned2" `
|
58 |
+
--train_batch_size=$train_batch_size `
|
59 |
+
--dataset_repeats=$dataset_repeats `
|
60 |
+
--learning_rate=$learning_rate `
|
61 |
+
--max_train_steps=$max_train_set `
|
62 |
+
--use_8bit_adam --xformers `
|
63 |
+
--mixed_precision=$mixed_precision `
|
64 |
+
--save_every_n_epochs=$save_every_n_epochs `
|
65 |
+
--train_text_encoder `
|
66 |
+
--save_precision="fp16"
|
67 |
+
|
68 |
+
# accelerate launch --num_cpu_threads_per_process $num_cpu_threads_per_process D:\kohya_ss\diffusers_fine_tuning\fine_tune_v1-ber.py `
|
69 |
+
# --pretrained_model_name_or_path=$train_dir"\fine_tuned\last.ckpt" `
|
70 |
+
# --in_json $train_dir"\meta_lat.json" `
|
71 |
+
# --train_data_dir=$train_dir"\"$training_folder `
|
72 |
+
# --output_dir=$train_dir"\fine_tuned2" `
|
73 |
+
# --train_batch_size=$train_batch_size `
|
74 |
+
# --dataset_repeats=$([Math]::Ceiling($dataset_repeats / 2)) `
|
75 |
+
# --learning_rate=$learning_rate `
|
76 |
+
# --max_train_steps=$([Math]::Ceiling($max_train_set / 2)) `
|
77 |
+
# --use_8bit_adam --xformers `
|
78 |
+
# --mixed_precision=$mixed_precision `
|
79 |
+
# --save_every_n_epochs=$save_every_n_epochs `
|
80 |
+
# --save_half
|