Hugging Face's logo

Bedovyy
/
arcaillous-nbxl-v10

Text-to-Image
Diffusers
Safetensors
English
StableDiffusionXLPipeline
Inference Endpoints
Model card Files Community
1
arcaillous-nbxl-v10
File size: 3,307 Bytes 
aa9c3cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
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
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
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
 
---
license: other
license_name: fair-ai-public-license-1.0-sd
license_link: https://freedevproject.org/faipl-1.0-sd/
language:
- en
library_name: diffusers
pipeline_tag: text-to-image
base_model: Laxhar/noobai-XL-1.0
datasets:
- pls2000/pixiv20161029_20241026_monthly_rank_1_50
- pls2000/aiart_channel_nai3_geachu
- cagliostrolab/860k-ordered-tags
---

## Training

Trained in 2steps, `Lion8bit` for quick training and `Lion` for detail.

- Tool: kohya-ss/sd-scripts
- GPUs: 2x RTX3090


### arcaillous-nbxl-v10b.safetensors
```
NCCL_P2P_DISABLE=1 NCCL_IB_DISABLE=1 accelerate launch --num_cpu_threads_per_process 8 sdxl_train.py $@ \
        --pretrained_model_name_or_path="/ai/data/sd/models/Stable-diffusion/noobaiXLNAIXL_epsilonPred10Version.safetensors" \
        --dataset_config="arca_nbxl.toml" \
        --output_dir="results/ckpt" --output_name="arcaillous-nbxl-v10b" \
        --save_model_as="safetensors" \
        --train_batch_size 4 --gradient_accumulation_steps 64 \
        --learning_rate=1e-5 --optimizer_type="Lion8bit" \
        --lr_scheduler="constant_with_warmup" --lr_warmup_steps 100 --optimizer_args "weight_decay=0.01" "betas=0.9,0.95" --min_snr_gamma 5 \
        --sdpa \
        --no_half_vae \
        --cache_latents --cache_latents_to_disk \
        --gradient_checkpointing \
        --full_bf16 --mixed_precision="bf16" --save_precision="fp16" \
        --ddp_timeout=10000000 \
        --max_train_epochs 4 --save_every_n_epochs 1 \
        --log_with wandb --log_tracker_name kohya-ss --wandb_run_name "arca_nbxl_`date +%y%m%d-%H%M`" --logging_dir wandb
```

### arcaillous-nbxl-v10.safetensors
```
NCCL_P2P_DISABLE=1 NCCL_IB_DISABLE=1 accelerate launch --num_cpu_threads_per_process 8 sdxl_train.py $@ \
        --pretrained_model_name_or_path="/ai/train/ckpt/arcaillous-nbxl-v10b.safetensors" \
        --dataset_config="arca_nbxl.toml" \
        --output_dir="results/ckpt" --output_name="arcaillous-nbxl-v10" \
        --save_model_as="safetensors" \
        --train_batch_size 1 --gradient_accumulation_steps 256 \
        --learning_rate=1e-5 --optimizer_type="Lion" \
        --lr_scheduler="constant_with_warmup" --lr_warmup_steps 100 --optimizer_args "weight_decay=0.01" "betas=0.9,0.95" \
        --min_snr_gamma 5 --ip_noise_gamma 0.05 --debiased_estimation_loss \
        --xformers \
        --no_half_vae \
        --cache_latents --cache_latents_to_disk \
        --gradient_checkpointing \
        --full_bf16 --mixed_precision="bf16" --save_precision="fp16" \
        --ddp_timeout=10000000 \
        --max_train_epochs 8 --save_every_n_epochs 1 --save_every_n_steps 200 \
        --log_with wandb --log_tracker_name kohya-ss --wandb_run_name "arca_nbxl_`date +%y%m%d-%H%M`" --logging_dir wandb
```

### arca_nbxl.toml
```
[general]
shuffle_caption = true
caption_tag_dropout_rate = 0.2
keep_tokens_separator = "|||"
caption_extension = ".txt"

[[datasets]]
  enable_bucket = true
  min_bucket_reso = 512
  max_bucket_reso = 4096
  resolution = 1024

  [[datasets.subsets]]
  image_dir = "/ai/data/sd/datasets/danbooru-gs"
  num_repeats = 1
  [[datasets.subsets]]
  image_dir = "/storage/pls2000_pixiv20161029_20241026_monthly_rank_1_50/to_train"
  num_repeats = 1
  [[datasets.subsets]]
  image_dir = "/storage/aichan/to_train"
  num_repeats = 1
```