A style LoRA trained on Y2K revival and K-pop-inspired female streetwear fashion. Designed to capture midriff-baring outfits, low-rise baggy pants, jackets, belts, beanies and trendy urban looks with both candid and studio-style vibes.
Great for: fashion shoots, character design, street-style portraits, urban anime aesthetics, or anything with early 2000s style or recent 2024/5 style.
Trained on a mix of angles (front-facing, from behind, three-quarter view, etc.) and a mix of poses (standing, bent over, sitting, walking) with emphasis on lots of midriff and pelvis (if you want that) with a healthy dose of subtle attitude.
Recommended settings:
Trigger word: y2kstr33twear
Weight: 0.2 - 0.4 (tested in ComfyUI, A1111 may need higher, but start with 0.3)
Works well with: realistic, photographic or anime models (examples are generated using a range of sd1.5 models to show the diversity)
Pair with tags like: crop top, long sleeves, cargo pants, baggy pants, hip hop pants, jacket, beanie, hat, pelvis, butt crack, candid, anime, trainers, front-facing, from behind, side view, three-quarter view, standing, bent over, sitting, walking
Description
Initial version
Trained using 90 images of 2k or higher and manually created tokens
style settings: batch 4, unet_lr = 0.0001, text_encoder_lr = 5e-5, network_dim = 64, network_alpha = 48, repeats 3, epochs 8 (7th best fit)
images scaled to 768 for full run
images scaled to 512 for second run
resulting LoRAs merged at 1.0 1.0