title: Image Restoration: Task and Fidelity Guide description: Image restoration must declare the degradation and fidelity target; choose a task-compatible deterministic or diffusion method, then validate measured recovery separately from plausible but invented detail. category: techniques tags: [image-restoration, denoising, deblurring, super-resolution, jpeg-artifacts, diffusion, fidelity, evaluation] aliases: ["Image Restoration Survey", "Image Enhancement", "Denoising Survey"]
Image Restoration: Task and Fidelity Guide¶
Image restoration needs a declared degradation and a declared fidelity target. “Looks better” is not sufficient for product, documentary, scientific, or forensic material: a generative method can produce plausible detail that was not supported by the observed image.
Start with the restoration contract¶
Define:
- degradation type and known acquisition history, such as noise, blur, compression, missing area, low light, or limited resolution;
- required output behavior: measurement-faithful recovery, perceptual cleanup, or clearly labeled creative reconstruction;
- available reference target or a source-specific validation method;
- protected content such as text, faces, product geometry, colors, and marks; and
- allowed model/data/license boundary.
Do not combine degradation classes under one generic “enhance” action unless the model was trained and validated for that combined contract.
Method families¶
Task-specific deterministic models¶
Task-specific CNN and transformer approaches remain appropriate when the degradation is known and pixel-faithful recovery matters. Examples include SwinIR, NAFNet, and Restormer. Their published results apply to particular tasks, datasets, and checkpoints; select a compatible release instead of treating any of them as a universal restoration model.
Conditional diffusion restoration¶
Palette demonstrates conditional diffusion across image-to-image tasks including colorization, inpainting, uncropping, and JPEG restoration. Diffusion methods can model a broader output distribution, but that flexibility makes unsupported detail a central acceptance risk.
IR-SDE is another research formulation that uses a mean-reverting stochastic differential equation for restoration. It does not establish a default schedule or a production configuration for a different checkpoint.
Editing and foundation-model workflows¶
General image-editing systems may be useful for a bounded, mask- and instruction-controlled repair. They need source-preservation evaluation and must be kept distinct from metric-oriented denoising or super-resolution. See RealRestorer for one release-specific workflow rather than a ranking over all restorers.
Evaluation that catches invention¶
Use a validation set with known targets where possible, and report the metric in its task context. Add visual review for:
- hallucinated text, geometry, texture, or small-object detail;
- color and tone drift, especially after compression or low-light correction;
- seams and discontinuities at masks, crops, or tiles;
- preservation of product, document, or medical/scientific evidence; and
- behavior on degraded inputs outside the training distribution.
Separate a metric result from a release decision. A high reconstruction metric does not automatically prove factual preservation, and a visually attractive diffusion output may be unacceptable when pixels are evidence.
Release record¶
Store the input digest, degradation diagnosis, model/checkpoint, parameters, reference target or validation fixture, output, acceptance result, and an explicit label for any creative reconstruction. Re-run validation after a model, runtime, or input-domain change.