---

name: image-training description: "Choose, edit, validate, and launch KAIR image training configurations for denoising, super-resolution, blind SR, deblocking, PSNR, GAN, DataParallel, and DDP workflows." metadata: disco-role: operating disable-model-invocation: true license: MIT

KAIR image-training

Use this sub-skill when the task is to train or resume KAIR image restoration models, adapt KAIR training option JSON files, choose a training entry script, or diagnose training configuration failures for image workflows.

This sub-skill covers KAIR image training for DnCNN, FDnCNN, FFDNet, DRUNet, USRNet, SRMD, DPSR, MSRResNet, RRDB/RRDBNet, IMDN, BSRGAN-style blind SR, SwinIR image SR/denoising/JPEG deblocking, PSNR-oriented training, GAN training, DataParallel, and DistributedDataParallel.

Do not use this sub-skill for:

• image inference or metrics; route to ../image-testing/SKILL.md.
• VRT/RVRT video restoration; route to ../video-restoration/SKILL.md.
• dataset splitting, LMDB creation, video regrouping, or folder-layout repair; route to ../data-preparation/SKILL.md.

Start here

1. Identify the task family and use references/training-workflows.md to choose the KAIR command entry point and option JSON template.
2. Edit a copied option JSON using references/configuration-reference.md; do not treat the bundled examples as exhaustive source configs.
3. Run scripts/validate_training_config.py --config <option-json> before launching training. It is read-only and does not import KAIR.
4. For multi-GPU runs, ensure gpu_ids, visible devices, --nproc_per_node, and --dist True agree before spending training time.
5. For resume or fine-tuning, read the checkpoint behavior in references/training-workflows.md before assuming path.pretrained_netG is honored directly.
6. If the failure is an option parse, path, selector, DDP, CUDA, checkpoint, or OOM issue, use references/troubleshooting.md.

Training commands are intentionally provided as templates to run in the user's own KAIR checkout. Full training is expensive and is not a safe smoke test.