BAXLv3 is up now!
Focus on generalization ability, cover art style of most copyright characters without catastrophic forgetting / overfitting.
What you should expected from this update:
Better generalization
Closer to Blue Archive style
Beautiful glare
Much improved LoRA compatibility based on KohakuΔ.
What you should NOT expected from this update:
Better anatomy, always good hands and feet
easy to use
focus on given prompts (compare to previous version)
multiple person / nsfw support (this is not gonna get any close to pony.)
detailed background
How-to:
Prompting logic is same to KohakuΔ.
1girl, <character>, <general tag>, <quality tag>
Negative:
lowres, error, worst quality, low quality, jpeg artifacts, watermark, signature, username
Keep prompt clean and tidy. Suggest prompt length < 75 tokens.
Suggetion:
CFG: 4-7 （higher will have slightly better anatomy）
Sampler: Euler A @ 25steps
Resolution: 768 - 1792(ultrawide only), +32 each step.
Use DanTagGen.
Set a low temperature (<1)
Ban words: sketch, comic, flat color, .*official.*, .*boy.*, mecha, no humans, text, pixel art, speech bubble
Total tag length: short
Use Hires.Fix
Upscaler: DATx2
Denoising: 0.4-0.5
Use ADetailer.
face_yolov8n.pt
If you want generate typical color, put color scheme to at first.
What's difference:
Base model switch to KohakuΔ rev1.
Reason
Aesthetic rating which AniXL used is NOT friendly to Celluloid art style, quality / negative tags of AniXL are effortless, sometimes even drawback for drawing clean, thin lines. Also it seems like AniXL has some overfitted characters.
👆TBH, I haven't use same training methods on AniXL3.1 yet. Maybe it can be fixed by the final training config I use.
Meanwhile, KohakuΔ is undertrained, which means flexible, more friendly to fine-tune art style.
I have no idea how to tag like AniXL3.1, but I know how to tag like KohakuΔ, since I could get the exact same dataset from Hakubooru.
Better generalization & horizonal composition.
KohakuΔ rev1 is not good at anatomy & composition, due to lack of training time (dual 3090 bro, what you expected).
Due to Regularization, this fine-tune do make things better (especially horizonal image), but I'm not gonna lie you, BAXLv3-Δ is VERY EASY to generate bad hand / extra legs, get worse if the prompts have problems (too long / semantic repetition / immoderate tag order).
Regularization / Datasets Update.
Regularization Danbooru Dataset, class token=solo:
1000+ horizonal image
fav count > 30
tag with "1girl, solo"
1000+ vertical image
fav count > 30
tag with "1girl, solo"
artist who is celluloid art style & BA art style imitator
Dataset update to 0068 events CG, include 5th PV screenshots.
Training Details:
Image count: 574 without repeat.
min_bucket_reso = 256
max_bucket_reso = 4096
bucket_reso_steps = 32
train_batch_size = 2
gradient_accumulation_steps = 32
learning_rate = 7.5e-6 Unet only
lr_scheduler = "constant_with_warmup"
lr_warmup_steps = 100
optimizer_type = "Lion8bit"
min_snr_gamma = 5
Batch size = 2
mixed_precision = "fp16"
full_fp16 = true
optimizer_args = [ "weight_decay=0.1", "betas=0.9,0.95" ]
shuffle_caption = true
weighted_captions = false
keep_tokens = 0
caption_tag_dropout_rate = 0.1
Known issues:
halo / heterochromia / .*focus.* seems to be overfitted in some circumstances. Can be helped by put them in negative tags.
Background is not stable, keen to draw something unnecessary even if you type simple background, white background
License & Disclaimer
License:
Dislcaimer:
According to Blue Archive's Fanart Guidelines ( JP | CN ), this model shall not be used for any kind of commercial use, including but not limited to selling this model or merges of this model, selling image generated by this model or merges of this model, or 'paid member exclusive' for monetization platforms like Patreon/Fanbox/etc.
