This is a SDXL LORA for my character Beata. Learned quite a lot creating this character and didn't overcook her lora like my other one. Used 60 images this time, with a lot different angles, poses, shots, lighting. Kept the description tags out of generation so she's got flexible hairstyles, eye colors, outfits, etc. Really tested this one for versatility.
Suggested Strength:
0.6-0.9 Lower is more flexible, higher will start aligning with training images more.
Trigger words:
beata
Suggested Prompts:
set eye color, hair color, hair style, and descriptive outfit
example:
(beata), makeup, brown eyes, dark brown twin braids framing her face,
small waist, toned athletic body, , etc.
Description
Dataset
Folder name: 10_charactername or 8_charactername (repeats)
Repeats: 8–10
Epochs: 1
Max Train Epoch: 10-20
→ This gives you ~2000–3000 total steps (ideal range for characters). Will be divided by Train Batch Size
LoRA Tab – Basic
Pretrained Model: JuggernautXL (or latest version)
LoRA Type: Standard
Network Rank (Dim): 64
Network Alpha: 32
Train Batch Size: 2
Gradient Accumulation Steps: 2–4 (effective batch 4–8)
Mixed Precision: bf16
Resolution: 1024×1024
Enable Bucketing: Yes
Save Every N Epochs: 1
Advanced / Optimizer
Optimizer: AdamW8bit
Learning Rate:
UNet: 1e-4 (0.0001)
Text Encoder: 5e-5
LR Scheduler: cosine_with_restarts
LR Scheduler Num Cycles: 3
Min SNR Gamma: 5
Cache Latents: Yes
Gradient Checkpointing: Off (no need with 32GB)
Color Augmentation: OFF
Shuffle Captions: Yes
Keep Tokens: 1 (essential)
Flip Augmentation: Off
Captioning
Trigger word first in every .txt (e.g. sks_person, or ohwx_john,).
Use WD14 Tagger → then add prefix via Utilities.
Keep detailed natural descriptions.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.