Denoising Architectures: Degradation and Evidence Contract¶
Architecture names do not identify the degradation to remove or the detail to preserve. A denoiser may estimate sensor noise, compression artifacts, blur, rain, a synthetic perturbation, or a mixture; each task needs a separate input, fidelity, and evaluation contract.
The NTIRE 2025 Image Denoising Challenge report reports methods and scores for its specified AWGN sigma-50 challenge. Its findings are useful evidence for that benchmark. They do not rank architectures universally across camera pipelines, noise mixtures, artistic retouching, forensic recovery, runtime limits, or protected-detail requirements.
Declare the restoration target¶
For each model evaluation or deployment, record:
- source asset/capture provenance, orientation, color profile, dynamic range, compression history, alpha/mask policy, and declared degradation model;
- model/checkpoint, code/runtime/version, compatible input range, color domain, preprocessing, padding, crop/tiling/blend policy, and output transform;
- authorized paired or normal-reference data, annotation/ground-truth policy, group-disjoint splits, and any synthetic degradation generator;
- task metrics and preservation metrics, evaluated on a source-disjoint set; and
- visual review of noise removal, texture, text, edges, product geometry, faces, shadows, gradients, color, and protected regions at delivery size.
Changing the camera, codec, noise distribution, crop, working color space, tile plan, or checkpoint creates a new configuration. A score from a different degradation or source split is not evidence for it.
Choose by measured failure behavior¶
Convolutional, transformer, state-space, Fourier, and diffusion-based approaches make different locality, conditioning, and generation assumptions. Use the documented release only where its assumptions match the declared task, then compare it against a named baseline under identical inputs and export path.
Measure latency and peak memory separately from recovery quality. If a method generates plausible texture or detail that is not supported by the input, label that output as derived/generated rather than recovered source signal. For masked correction, record the mask and separately verify unchanged regions; an apparent no-op outside a mask is not proven without comparison.
Failure boundary¶
If degradation assumptions, model compatibility, ground-truth authority, source-disjoint evaluation, or protected-detail review is absent, return a visible review/failure state. Do not select a model because it tops another leaderboard, silently switch to a generative prior, or represent plausible detail as factual restoration.