Hollowlight - Manga & Anime Ink Style | SDXL LoRA

All images made with NoobAI-XL V-Pred 1.0 + Hollowlight v3.0 at strength 1.0, Euler / normal, 28 steps, CFG 5, RescaleCFG 0.4, with face and hand detailers.

🚨 V3.0 UPDATE NOTES (READ BEFORE DOWNLOADING)

Welcome to V3.0! Resumed from the V2.0 foundation (Step 2250) and trained further on an expanded dataset focused heavily on 2-character interactions, full-body framing, and complex compositions.

Just like V2, I am releasing two checkpoints for V3.0 so you can pick your preferred finish:

• Step 0800 (Main Download File — Recommended): The true sweet spot. Keeps the softer, natural, hand-drawn G-pen shading of V2's main checkpoint while holding complex 2-character and full-body compositions together without over-sharpening.

• Step 0900 (Alternative Download File): The sharper, punchier version. Features tighter line contrast, crisp rim-lighting separation, and denser micro-details in hair and complex backgrounds if you want maximum sharpness.

What's New in V3.0:

• 2-Character & Multi-Subject Stability: Much cleaner spacing, interaction, and anatomy when generating two characters in the same frame.

• Stronger Full-Body & Perspective Compositions: Holds wide shots, dynamic angles, and multi-layered backgrounds together without breaking structure.

• Hand Detailer Recommended: Just like faces in wide shots, using a hand detailer pass is recommended for complex full-body poses.

📦 V2.0 UPDATE NOTES (ARCHIVE)

• Step 2250 (V2 Main): Features a softer, less sharpened look with a bit more flexibility. Great if you want to avoid that super-crisp "AI sharpness" look and want the lines to feel more natural and hand-drawn.

• Step 3000 (V2 Alternative): Features maximum texture sharpness, crisper hair strands, and higher overall detail contrast.

• What Was New in V2: Retrained on 272 black-and-white manga-style images for drastically improved composition layout.

⚠️ IMPORTANT: V-PRED Settings Required

This version is a V-PRED LoRA. You MUST use it with NoobAI-XL V-Pred 1.0 as your base model, and ensure you use v-prediction sampling with Zero Terminal SNR turned ON. If you do not change these settings in your UI, your generation outputs will look completely burnt out or washed out.

📖 General Model Manual & Guide

Hollowlight focuses on expressive character artwork, clean ink lines, strong black-and-white contrast, and manga-inspired illustrations.

• Trigger word: u616 (Start your prompts with this)

• Base model: NoobAI-XL V-Pred 1.0

⚙️ Recommended Settings:

• LoRA strength: 1.0

• CLIP strength: 1.0

• Sampling steps: 28

• CFG: 5 or 7 (Using RescaleCFG 0.4 helps significantly)

• Sampler: Euler

• Scheduler: Normal

• Resolution: Portrait: 832 x 1216, Square: 1024 x 1024, Landscape: 1216 x 832

• Face & Hand Detailers Recommended: (See pro-tip below)

💡 Pro-Tip for Better Faces & Hands (Wide / Full-Body Shots):

In wide, 2-character, or full-body shots, faces and hands can sometimes come out small and simple due to resolution constraints. Running a detailer pass completely fixes this!

• Recommended Setup: Impact Pack FaceDetailer + Impact Subpack, with face_yolov8m.pt (from Bingsu/adetailer on Hugging Face) in models/ultralytics/bbox (close-ups work fine without it).

• Settings: guide_size 640, denoise 0.6, bbox_threshold 0.1, steps 28, CFG 5, Euler / normal. Leave the detailer prompt empty so it uses your main prompt.

• Troubleshooting: If the detailer seems to do nothing (the image comes back unchanged), the detector isn't finding the small face or hand in your wide shot—lower bbox_threshold toward 0.1.

Enjoy experimenting with Hollowlight v3.0!