Hollowlight — Manga Ink Backgrounds & Scenery (NoobAI V-Pred)

Part of the Hollowlight dataset series. This LoRA generates traditional black-and-white manga backgrounds featuring crisp ink linework, detailed crosshatching, and high-contrast foliage.

Important Base Model Note: This LoRA was trained on NoobAI (V-Pred). The embedded file metadata reads sdxl_base_v1-0 due to Kohya's default export label, please make sure you run this on a NoobAI V-Pred checkpoint for proper results.

Usage & Prompting

• Trigger Word: u617

• Recommended Weight: 1.0 (Tested exclusively at 1.0)

• Helpful Tags: scenery, no humans, monochrome, greyscale, manga, traditional media, ink, crosshatching, forest, river, mountains, trees, grass, rocks, clouds

Strengths & Limitations

• What it's great at: Natural outdoor environments—forests, rivers, rolling hills, dense trees, tall grass, and manga-style clouds.

• Current limitations: complex architecture are weaker and may come out distorted, as the v1 dataset contained very few architectural images. A future update is planned to add more buildings and structures.

Training Specs

• Base Model: NoobAI (V-Pred)

• Dataset: 101 images | 1,500 steps | 30 epochs

• Network: Rank 32 / Alpha 16

• Learning Rate: Unet LR 6e-5 (Text Encoder not trained)

• Optimizer & Scheduler: AdamW8bit, constant_with_warmup (50 warmup steps)

• SNR Settings: Min SNR gamma 5, Zero Terminal SNR enabled

Upscaling: if aw outputs look too soft for u, upscale in pixel space with an ESRGAN-type model instead of a latent upscale. I use 4x-AnimeSharp (4x, then downscale 50% with Lanczos) or 2x-AnimeSharpV4.

Latent upscaling (hires fix with nearest-exact or bislerp) gave mesh and grid artifacts on the line work, so skip it.