
Hui Buh from German trash movies Hui Buh: The Goofy Ghost and Hui Buh and the Witchs Castle. The character design is rougly based on the appearence of Michael “Bully” Herbig.
Created as bf16 DoRa, so it might require a fairly modern inference engine and/or GPU.
Tags
? Primary tag
HuiBuh
? Situational general tags
1boy, solo
? Essential face tags
bob cut, dark brown hair
? Situational face tags
blunt bangs, thick eyebrows
? Essential clothing tags
metal collar, armor
? Situational clothing tags
chain, breastplate
Inference
Recommended checkpoints:
Recommended settings:
-
Sampler: Euler a
-
Steps: 40
-
CFG scale: 7
-
Clip skip: 1 (yes, really)
Changelog
v1.0.11 → v1.2.25
-
Decreased steps from 1040 to 832
-
Increased network alpha from 12.5% dim to 25% dim
-
Increased batch size from 1 to 4
v1.2.25 → v1.2.35
-
Trained for 10 more epochs
v1.2.35 → v1.2.35-lite
-
Reduced network dim from 32 to 8
-
Increased network alpha from 25% dim to 50% dim
v1.2.35 → v1.3.35
-
Decreased batch size from 4 to 1
v1.3.35 → v1.4
-
Changed learning rate scheduler from “cosine” to “constant”
-
Reduced learning rate from 0.001 to 0.0005
-
Increased steps from 832 to 1248
-
Disabled tag shuffling
Development
This is my first ever lora developed, so please expect some training mistakes. I created the dataset from movie screencaps. It’s publicly available here. The training was done on a Colab L4 VM (22.5 gb VRAM) using custom software based on kohya-ss/sd-scripts.
Kohya invocation:
accelerate launch /content/repos/kohya-ss/sdxl_train_network.py --pretrained_model_name_or_path /content/inputModel/checkpoint.safetensors --vae /content/inputModel/vae.safetensors --train_data_dir dataset/train/dreambooth --output_dir /content/model --logging_dir /content/log --resolution 1024,1024 --save_model_as safetensors --text_encoder_lr 0.0001 --unet_lr 0.001 --network_dim 32 --network_alpha 8 --output_name hui_buh --max_train_steps 832 --caption_extension .txt --cache_latents --train_batch_size 4 --seed 1910506442 --shuffle_caption --caption_separator ', ' --cache_info --no_half_vae --network_module lycoris.kohya --network_args preset=full algo=lora dora_wd=True --optimizer_type adafactor --optimizer_args scale_parameter=False relative_step=False warmup_init=False --save_every_n_epochs 1 --learning_rate 0.001 --lr_scheduler cosine --highvram --xformers --gradient_checkpointing --fused_backward_pass --mixed_precision no --save_precision bf16 --sample_sampler euler_a --sample_prompts prompt.txt --sample_at_first --sample_every_n_epochs 1 --log_with wandb --wandb_api_key ? --wandb_run_name hui_buh-1723892649 --log_config --huggingface_repo_id Jaidchen/sdxl-pony-lora-hui-buh --huggingface_path_in_repo model --huggingface_repo_type model --huggingface_token ? --save_state_to_huggingface --async_upload --metadata_title 'Hui Buh' --metadata_author Jaidchen
Collaboration
Created in collaboration with:
-
Canas and Volke by MadlyLaughing
-
Aadit and Hisao Nakai by NorthSpirit
-
Ferdinand von Aegir and Hanneman by bsinky
-
Stark and Jack Atlas by ChameleonAI
-
Space Dandy and Bobobo by Ravenfoot
-
Kim Dokja, Yoo Joonghyuk and Min Ryu by Shufflebox
Thanks to mfcg for collecting the links!
描述:
训练词语: HuiBuh
名称: hui_buh-v1.2.25.safetensors
大小 (KB): 252581
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success