OpenComic AI ships 15 pre-trained model weights covering three image-restoration tasks — artifact removal, descreening, and upscaling — each available in three neural-network architectures. All weights are distributed in NCNN format and are ready to load without any post-processing step.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.
Architectures
Each task is covered by three architecture tiers so you can balance quality against compute budget:- ESRGAN (full): the highest-quality variant, using the full ESRGAN backbone with 23 residual dense blocks. Delivers the best restoration fidelity but requires the most VRAM and is the slowest at inference time.
- ESRGAN Lite: a slimmed-down ESRGAN variant (32 filters, 12 blocks) that trades a small amount of quality for significantly faster inference and lower memory usage. A good default for most desktop hardware.
- Compact: the smallest and fastest architecture, designed for real-time use or resource-constrained environments such as integrated graphics or lower-end GPUs. Trained for 450,000 iterations versus 1,000,000 for ESRGAN variants.
Where to Find the Weights
Model weights are available from three locations:models/folder in this repository — weights are bundled directly alongside the training configuration files.- Releases page — each tagged release on GitHub bundles the corresponding NCNN
.bin+.paramfile pairs. - ollm/opencomic-ai-models — a dedicated companion repository that hosts all current and historical weights.
Pre-Training Chain
The training pipeline is sequential by design. Artifact-removal models are trained first from scratch, then their weights are used as the starting point for both the descreen and upscale-2x models. The upscale-3x and upscale-4x models in turn start from the upscale-2x checkpoint. This chain means each specialist model inherits broad low-level restoration knowledge before being fine-tuned on its specific task.All Models
| Model Name | Architecture | Task | Scale | Iterations |
|---|---|---|---|---|
opencomic-ai-artifact-removal-compact | Compact | Artifact Removal | 1× | 450,000 |
opencomic-ai-artifact-removal-lite | ESRGAN Lite | Artifact Removal | 1× | 1,000,000 |
opencomic-ai-artifact-removal | ESRGAN | Artifact Removal | 1× | 1,000,000 |
opencomic-ai-descreen-hard-compact | Compact | Descreen | 1× | 450,000 |
opencomic-ai-descreen-hard-lite | ESRGAN Lite | Descreen | 1× | 1,000,000 |
opencomic-ai-descreen-hard | ESRGAN | Descreen | 1× | 1,000,000 |
opencomic-ai-upscale-2x-compact | Compact | Upscale | 2× | 450,000 |
opencomic-ai-upscale-2x-lite | ESRGAN Lite | Upscale | 2× | 1,000,000 |
opencomic-ai-upscale-2x | ESRGAN | Upscale | 2× | 1,000,000 |
opencomic-ai-upscale-3x-compact | Compact | Upscale | 3× | 450,000 |
opencomic-ai-upscale-3x-lite | ESRGAN Lite | Upscale | 3× | 500,000 |
opencomic-ai-upscale-3x | ESRGAN | Upscale | 3× | 500,000 |
opencomic-ai-upscale-4x-compact | Compact | Upscale | 4× | 450,000 |
opencomic-ai-upscale-4x-lite | ESRGAN Lite | Upscale | 4× | 500,000 |
opencomic-ai-upscale-4x | ESRGAN | Upscale | 4× | 500,000 |
OpenComic AI model weights are licensed under CC BY 4.0 (Creative Commons Attribution 4.0 International). You are free to use, redistribute, and adapt the weights — including for commercial purposes — as long as you provide appropriate credit. The recommended attribution format is:
OpenComic AI by Oleguer Llopart Mora (CC BY 4.0)When redistributing weights, include
https://creativecommons.org/licenses/by/4.0/
LICENSE_MODELS or equivalent attribution information in your package or release notes. The source code in this repository is separately licensed under the MIT License.Explore by Task
Artifact Removal
Remove JPEG, WebP, AVIF, and JXL compression noise from scanned comic pages. Trained on 400,000 image pairs and used as the pretrain base for all other models.
Descreen
Eliminate halftone dot and line patterns introduced by the printing process while preserving linework and tonal gradients. Trained on 120,000 image pairs.
Upscale
Increase resolution 2×, 3×, or 4× with sharp linework and clean rendering. Nine models covering all three scale factors across all three architectures.