Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/ollm/opencomic-ai-training/llms.txt

Use this file to discover all available pages before exploring further.

Printed comics are rendered using halftone screening — a process where continuous tones are broken into grids of tiny dots or lines so that they can be reproduced with a fixed-colour ink. When you scan a printed page, those dot patterns are captured by the scanner and appear as a visible repeating texture in the digital image. Descreening removes this printing artefact while reconstructing the smooth tonal gradients and clean linework that existed in the original artwork. Unlike a simple blur, the descreen models are trained to distinguish halftone structure from real image detail and to leave edges, hatching, and fine lines untouched.

Dataset

Each descreen model was trained on 120,000 image pairs with 10 degraded variants per clean image. The training data combines four dataset splits to expose the model to the full variety of halftone patterns it will encounter in real scans:
  • opencomic-ai-descreen-hard — hard dots at a fixed 45° angle
  • opencomic-ai-descreen-hard-moire — hard dots plus moiré interference patterns
  • opencomic-ai-descreen-hard-any-angle — hard dots at random angles
  • opencomic-ai-descreen-hard-moire-any-angle — hard dots at random angles with moiré
Halftone types covered by the training data:
  • Hard dot halftones at standard print angles (45°)
  • Hard dot halftones at arbitrary angles (full rotation range)
  • CMYK multi-channel halftones (added in v2.0)

Model Variants

All three variants operate at 1× scale and produce NCNN weight files as a .bin + .param pair.
ModelArchitectureNCNN PathIterations
opencomic-ai-descreen-hardESRGANmodels/descreen/ncnn/1,000,000
opencomic-ai-descreen-hard-liteESRGAN Litemodels/descreen/ncnn/1,000,000
opencomic-ai-descreen-hard-compactCompactmodels/descreen/ncnn/450,000
models/descreen/ncnn/
  opencomic-ai-descreen-hard.bin
  opencomic-ai-descreen-hard.param
  opencomic-ai-descreen-hard-lite.bin
  opencomic-ai-descreen-hard-lite.param
  opencomic-ai-descreen-hard-compact.bin
  opencomic-ai-descreen-hard-compact.param

Pre-Trained From Artifact Removal

Each descreen model is fine-tuned from its corresponding artifact removal checkpoint rather than trained from scratch:
Descreen ModelPre-trained From
opencomic-ai-descreen-hardopencomic-ai-artifact-removal
opencomic-ai-descreen-hard-liteopencomic-ai-artifact-removal-lite
opencomic-ai-descreen-hard-compactopencomic-ai-artifact-removal-compact
Starting from a restoration-capable checkpoint gives the descreen models a strong prior for clean comic imagery before they are specialized on halftone removal.

Preserving Detail — v2.0 Improvements

OpenComic AI v2.0 made targeted improvements to prevent the descreen model from over-processing non-halftone content. The training dataset was extended to include circles and circles with a dot inside (such as eyes and irises) so the model learns to leave them untouched. Parallel lines and grids (such as ruled panels, crosshatching, and screen tones that are not halftones) were also added as negative examples. Additionally, a fix was applied to prevent the model from brightening very dark images or hard edges that happen to be adjacent to a halftone area.

Training Configuration

The traiNNer-redux training options for these models are located in:
options/train/ESRGAN/opencomic-ai-descreen-hard.yml
options/train/ESRGAN/opencomic-ai-descreen-hard-lite.yml
options/train/Compact/opencomic-ai-descreen-hard-compact.yml
The dataset generation presets used to produce the paired training data are:
options/opencomic-ai-descreen-hard.yml
options/opencomic-ai-descreen-hard-any-angle.yml
Generate the descreen training datasets with:
# Fixed-angle halftone dataset
npm run prepare && npm run generate -- \
  --options ./options/opencomic-ai-descreen-hard.yml \
  --krita ./krita-5.3.1-x86_64.AppImage

# Any-angle halftone dataset
npm run prepare && npm run generate -- \
  --options ./options/opencomic-ai-descreen-hard-any-angle.yml \
  --krita ./krita-5.3.1-x86_64.AppImage

Build docs developers (and LLMs) love