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.

Artifact removal models clean up the visual noise introduced when a comic image has been saved with lossy compression. Lossy codecs — particularly at lower quality settings — produce blocking (hard edges between 8×8 pixel regions), ringing (faint halos around sharp lines), and color banding (abrupt transitions in smooth gradients). The artifact removal models are trained to detect and suppress all of these degradations while leaving the underlying linework and color fills unchanged.

​
Dataset

Each model in this group was trained on 400,000 image pairs. Each clean source image was encoded with 100 distinct degradation variants, giving the model a broad exposure to the range of compression artefacts it will encounter in the wild. Degradations included in training:
  • JPEG at quality levels 30–100 (primary pass, probability 0.8)
  • WebP lossy compression (probability 0.1)
  • AVIF compression (probability 0.2; added in v2.0)
  • JXL (JPEG XL) compression (probability 0.2; added in v2.0)
  • Multiple stacked JPEG passes — up to four additional JPEG re-encodes applied in sequence (each at probability 0.1), simulating images that have been saved and re-saved multiple times
  • Small resize followed by re-upscale (probability 0.5, to simulate resampling artefacts)
  • Rotate (probability 0.2)
  • Mild blur (to simulate scan softening after compression)
The v2.0 dataset adds a small blur and resize step applied after the initial compression pass. This teaches the model to recognize artefacts that have been slightly softened — a common real-world scenario when scans are post-processed before distribution.

​
Model Variants

All three variants operate at 1× scale (restoration only, no upsampling) and produce NCNN weight files as a .bin + .param pair.
ModelArchitectureNCNN PathIterations
opencomic-ai-artifact-removalESRGANmodels/artifact-removal/ncnn/1,000,000
opencomic-ai-artifact-removal-liteESRGAN Litemodels/artifact-removal/ncnn/1,000,000
opencomic-ai-artifact-removal-compactCompactmodels/artifact-removal/ncnn/450,000
Each entry in models/artifact-removal/ncnn/ contains two files per model: 
models/artifact-removal/ncnn/
  opencomic-ai-artifact-removal.bin
  opencomic-ai-artifact-removal.param
  opencomic-ai-artifact-removal-lite.bin
  opencomic-ai-artifact-removal-lite.param
  opencomic-ai-artifact-removal-compact.bin
  opencomic-ai-artifact-removal-compact.param

​
Role as Pretrain Base

The artifact removal models are the foundation of the entire OpenComic AI training pipeline. Before the descreen and upscale models are trained, the artifact removal checkpoint is copied as their starting point. This means all downstream models inherit general restoration knowledge — edge sharpness, colour fidelity, noise suppression — before being fine-tuned on their specific task. 
opencomic-ai-artifact-removal        ──► opencomic-ai-descreen-hard
                                     ──► opencomic-ai-upscale-2x

opencomic-ai-artifact-removal-lite   ──► opencomic-ai-descreen-hard-lite
                                     ──► opencomic-ai-upscale-2x-lite

opencomic-ai-artifact-removal-compact ──► opencomic-ai-descreen-hard-compact
                                      ──► opencomic-ai-upscale-2x-compact

​
Training Configuration

The traiNNer-redux training options for these models are located in: 
options/train/ESRGAN/opencomic-ai-artifact-removal.yml
options/train/ESRGAN/opencomic-ai-artifact-removal-lite.yml
options/train/Compact/opencomic-ai-artifact-removal-compact.yml
The dataset generation preset used to produce the paired training data is: 
options/opencomic-ai-artifact-removal.yml
Pass this preset to the generation CLI to reproduce the training dataset: 
npm run prepare && npm run generate -- \
  --options ./options/opencomic-ai-artifact-removal.yml \
  --krita ./krita-5.3.1-x86_64.AppImage

Build docs developers (and LLMs) love

Get started for freeTalk to us