Disty0/sotediffusion-wuerstchen3-alpha2

SoteDiffusion Wuerstchen3

Anime finetune of Würstchen V3.
Currently is in early state in training.
No commercial use thanks to StabilityAI.

Release Notes

• Ran LLaVa on the images that has "english text" tag in it.
  This adds The text says "text" tag.
  If LLaVa has no idea what the text is, it describes the image instead.

UI Guide

SD.Next

URL: https://github.com/vladmandic/automatic/

Switch to the dev branch:

git checkout dev

Go to Models -> Huggingface and type Disty0/sotediffusion-wuerstchen3-alpha2-decoder into the model name and press download.
Load Disty0/sotediffusion-wuerstchen3-alpha2-decoder after the download process is complete.

Prompt:

very aesthetic, best quality, newest,

Negative Prompt:

very displeasing, worst quality, oldest, monochrome, sketch, realistic,

Parameters:
Sampler: Default

Steps: 30 or 40
Refiner Steps: 10

CFG: 8
Secondary CFG: 1 or 1.2

Resolution: 1024x1536, 2048x1152
Anything works as long as it's a multiply of 128.

ComfyUI

Please refer to CivitAI: https://civitai.com/models/353284

Code Example

pip install diffusers

import torch
from diffusers import StableCascadeCombinedPipeline

device = "cuda"
dtype = torch.bfloat16
model = "Disty0/sotediffusion-wuerstchen3-alpha2-decoder"

pipe = StableCascadeCombinedPipeline.from_pretrained(model, torch_dtype=dtype)

# send everything to the gpu:
pipe = pipe.to(device, dtype=dtype)
pipe.prior_pipe = pipe.prior_pipe.to(device, dtype=dtype)

# or enable model offload to save vram:
# pipe.enable_model_cpu_offload()



prompt = "1girl, solo, cowboy shot, straight hair, looking at viewer, hoodie, indoors, slight smile, casual, furniture, doorway, very aesthetic, best quality, newest,"
negative_prompt = "very displeasing, worst quality, oldest, monochrome, sketch, realistic,"

output = pipe(
    width=1024,
    height=1536,
    prompt=prompt,
    negative_prompt=negative_prompt,
    decoder_guidance_scale=1.0,
    prior_guidance_scale=8.0,
    prior_num_inference_steps=40,
    output_type="pil",
    num_inference_steps=10
).images[0]

## do something with the output image

Training Status:

GPU used for training: 1x AMD RX 7900 XTX 24GB
GPU Hours: 250 (Accumulative starting from alpha1)

dataset name	training done	remaining
newest	010	221
recent	010	162
mid	010	114
early	010	060
oldest	010	010
pixiv	010	032
visual novel cg	010	018
anime wallpaper	010	003
Total	88	620

Note: chunks starts from 0 and there are 8000 images per chunk

Dataset:

GPU used for captioning: 1x Intel ARC A770 16GB
GPU Hours: 350

Model used for captioning: SmilingWolf/wd-swinv2-tagger-v3
Command:

python /mnt/DataSSD/AI/Apps/kohya_ss/sd-scripts/finetune/tag_images_by_wd14_tagger.py --model_dir "/mnt/DataSSD/AI/models/wd14_tagger_model" --repo_id "SmilingWolf/wd-swinv2-tagger-v3" --recursive --remove_underscore --use_rating_tags --character_tags_first --character_tag_expand --append_tags --onnx --caption_separator ", " --general_threshold 0.35 --character_threshold 0.50 --batch_size 4 --caption_extension ".txt" ./

dataset name	total images	total chunk
newest	1.848.331	232
recent	1.380.630	173
mid	993.227	125
early	566.152	071
oldest	160.397	021
pixiv	343.614	043
visual novel cg	231.358	029
anime wallpaper	104.790	014
Total	5.628.499	708

Note:

• Smallest size is 1280x600 | 768.000 pixels
• Deduped based on image similarity using czkawka-cli

Tags:

Model is trained with random tag order but this is the order in the dataset if you are interested:

aesthetic tags, quality tags, date tags, custom tags, rating tags, character, series, rest of the tags

Date:

tag	date
newest	2022 to 2024
recent	2019 to 2021
mid	2015 to 2018
early	2011 to 2014
oldest	2005 to 2010

Aesthetic Tags:

Model used: shadowlilac/aesthetic-shadow-v2

score greater than	tag	count
0.90	extremely aesthetic	125.451
0.80	very aesthetic	887.382
0.70	aesthetic	1.049.857
0.50	slightly aesthetic	1.643.091
0.40	not displeasing	569.543
0.30	not aesthetic	445.188
0.20	slightly displeasing	341.424
0.10	displeasing	237.660
rest of them	very displeasing	328.712

Quality Tags:

Model used: https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/models/aes-B32-v0.pth

score greater than	tag	count
0.980	best quality	1.270.447
0.900	high quality	498.244
0.750	great quality	351.006
0.500	medium quality	366.448
0.250	normal quality	368.380
0.125	bad quality	279.050
0.025	low quality	538.958
rest of them	worst quality	1.955.966

Rating Tags

tag	count
general	1.416.451
sensitive	3.447.664
nsfw	427.459
explicit nsfw	336.925

Custom Tags:

dataset name	custom tag
image boards	date,
pixiv	art by Display_Name,
visual novel cg	Full_VN_Name (short_3_letter_name), visual novel cg,
anime wallpaper	date, anime wallpaper,

Training Parameters:

Software used: Kohya SD-Scripts with Stable Cascade branch
https://github.com/kohya-ss/sd-scripts/tree/stable-cascade

Base model: Disty0/sote-diffusion-cascade-alpha0

Command:

#!/bin/sh

CURRENT=$1
CURRENT_SUB=$2

PAST=$3
PAST_SUB=$4

LD_PRELOAD=/usr/lib/libtcmalloc.so.4 accelerate launch --mixed_precision fp16 --num_cpu_threads_per_process 1 stable_cascade_train_stage_c.py \
--mixed_precision fp16 \
--save_precision fp16 \
--full_fp16 \
--sdpa \
--gradient_checkpointing \
--train_text_encoder \
--resolution "1024,1024" \
--train_batch_size 2 \
--gradient_accumulation_steps 8 \
--learning_rate 1e-5 \
--learning_rate_te1 1e-5 \
--lr_scheduler constant_with_warmup \
--lr_warmup_steps 100 \
--optimizer_type adafactor \
--optimizer_args "scale_parameter=False" "relative_step=False" "warmup_init=False" \
--max_grad_norm 0 \
--token_warmup_min 1 \
--token_warmup_step 0 \
--shuffle_caption \
--caption_separator ", " \
--caption_dropout_rate 0 \
--caption_tag_dropout_rate 0 \
--caption_dropout_every_n_epochs 0 \
--dataset_repeats 1 \
--save_state \
--save_every_n_steps 256 \
--sample_every_n_steps 64 \
--max_token_length 225 \
--max_train_epochs 1 \
--caption_extension ".txt" \
--max_data_loader_n_workers 2 \
--persistent_data_loader_workers \
--enable_bucket \
--min_bucket_reso 256 \
--max_bucket_reso 4096 \
--bucket_reso_steps 64 \
--bucket_no_upscale \
--log_with tensorboard \
--output_name sotediffusion-wr3_3b \
--train_data_dir /mnt/DataSSD/AI/anime_image_dataset/combined/combined-$CURRENT/$CURRENT_SUB \
--in_json /mnt/DataSSD/AI/anime_image_dataset/combined/combined-$CURRENT/$CURRENT_SUB.json \
--output_dir /mnt/DataSSD/AI/SoteDiffusion/Wuerstchen3/sotediffusion-wr3_3b-$CURRENT/$CURRENT_SUB \
--logging_dir /mnt/DataSSD/AI/SoteDiffusion/Wuerstchen3/sotediffusion-wr3_3b-$CURRENT/$CURRENT_SUB/logs \
--resume /mnt/DataSSD/AI/SoteDiffusion/Wuerstchen3/sotediffusion-wr3_3b-$PAST/$PAST_SUB/sotediffusion-wr3_3b-state \
--stage_c_checkpoint_path /mnt/DataSSD/AI/SoteDiffusion/Wuerstchen3/sotediffusion-wr3_3b-$PAST/$PAST_SUB/sotediffusion-wr3_3b.safetensors \
--text_model_checkpoint_path /mnt/DataSSD/AI/SoteDiffusion/Wuerstchen3/sotediffusion-wr3_3b-$PAST/$PAST_SUB/sotediffusion-wr3_3b_text_model.safetensors \
--effnet_checkpoint_path /mnt/DataSSD/AI/models/wuerstchen3/effnet_encoder.safetensors \
--previewer_checkpoint_path /mnt/DataSSD/AI/models/wuerstchen3/previewer.safetensors \
--sample_prompts /mnt/DataSSD/AI/SoteDiffusion/Wuerstchen3/config/sotediffusion-prompt.txt

Limitations and Bias

Bias

• This model is intended for anime illustrations.
  Realistic capabilites are not tested at all.

Limitations

• Can fall back to realistic.
  Add "realistic" tag to the negatives when this happens.
• Far shot eyes can be bad.
• Anatomy and hands can be bad.
• Still in active training.

License

SoteDiffusion models falls under Fair AI Public License 1.0-SD license, which is compatible with Stable Diffusion models’ license. Key points:

1. Modification Sharing: If you modify SoteDiffusion models, you must share both your changes and the original license.
2. Source Code Accessibility: If your modified version is network-accessible, provide a way (like a download link) for others to get the source code. This applies to derived models too.
3. Distribution Terms: Any distribution must be under this license or another with similar rules.
4. Compliance: Non-compliance must be fixed within 30 days to avoid license termination, emphasizing transparency and adherence to open-source values.

Notes: Anything not covered by Fair AI license is inherited from Stability AI Non-Commercial license which is named as LICENSE_INHERIT. Meaning, still no commercial use of any kind.