Hollowlight - SDXL Manga Ink Style LoRA
A monochrome manga-style LoRA trained on 101 original black-and-white manga illustrations.
Hollowlight focuses on expressive character artwork, clean ink lines, strong black-and-white contrast, ands manga-inspired illustrations.
Trigger word: u616
Base model: NoobAI-XL V-Pred 1.0
Recommended settings:
LoRA strength: 1.0
CLIP strength: 1.0
Sampling steps: 28
CFG: 5
Sampler: Euler
Scheduler: Normal
Resolution: 832 x 1216
Training information:
Training images: 101
Network dimension: 16
Network alpha: 16
Selected checkpoint: 1500 steps
The 1500-step checkpoint was selected after comparing multiple checkpoints across different prompts and compositions.
Enjoy experimenting with Hollowlight!
Description
Hey everyone, welcome to V2! I couldn't decide between two amazing checkpoints, so I'm giving you both:
Step 2250: A softer, less sharpened look with a bit more flexibility. A good pick if you don't want the crisp "AI sharpness" look and want it more natural and hand drawn.
Step 3000: Features maximum texture sharpness, crisper hair strands, and higher overall detail contrast."
Trigger word: u616
Additionally Re Trained on 272 black-and-white manga-style images for better composition (NoobAI-XL V-Pred 1.0 base). TImportant: this is a V-PRED LoRA. Use it with NoobAI-XL V-Pred 1.0 and v-prediction sampling (zero terminal SNR on), or the output will look burnt or washed out.
Recommended settings:
- LoRA strength: 1.0
- Sampler: Euler, normal scheduler
- Steps: 28
- CFG: 5 (RescaleCFG 0.4 helps)
- Start prompts with: u616
Tip: in wide or full-body shots the face can come out small and simple. A face detailer pass (e.g. Impact Pack FaceDetailer, denoise ~0.5, guide size 512) fixes this.
Training: rank 32 / alpha 16, UNet only, AdamW8bit, LR 6e-5, batch size 2.