title: Paired Restoration Training: Evidence and Conditioning Contract description: "Paired restoration training learns a declared degraded-to-target mapping; it needs source-aligned and rights-cleared pairs, a model-compatible conditioning path, holdouts separated by source, and evaluation that distinguishes measured recovery from plausible invention." category: techniques tags: [image-to-image, paired-data, restoration, conditioning, provenance, evaluation] aliases: ["Image-to-Image Conditioning", "Paired Image Training", "Degradation Pipeline"]

Paired Restoration Training: Evidence and Conditioning Contract¶

Paired restoration training learns a declared mapping from a degraded input to an approved target. It is not ordinary text-to-image training with cleaner captions, and it must not present plausible generated texture as recovered source evidence.

Palette is a conditional image-to-image diffusion framework covering tasks such as colorization, inpainting, uncropping, and JPEG restoration. That paper supports the general need for an explicit image condition; it does not prescribe one channel layout, initialization, loss, scheduler, or training recipe for every modern model.

Define the pair¶

For every sample, retain:

• original/source authority and rights record;
• degraded input, target, and their hashes;
• alignment/crop/resize/color-space policy;
• declared degradation or approved retouch operation;
• editable versus protected/unknown regions; and
• reviewer decision and any uncertainty or exclusion reason.

Source and target must depict the same approved scene or a deliberately specified edit. If the target contains manual beautification, generative fill, or unknown historical changes, label that fact rather than treating it as a clean ground truth.

Conditioning must match the model¶

Image-conditioned systems may use channel concatenation, cross-attention, a dedicated control branch, an image encoder, or another architecture-specific mechanism. Use only the conditioning path documented and trained for the pinned checkpoint. Changing input channel counts, injecting latents, or copying a conditioning module from another framework creates a new model configuration that requires its own initialization, training, and evaluation.

Record the full conditioning contract:

• base model/checkpoint and VAE/encoder revisions;
• runtime, scheduler/objective, and conditioning module implementation;
• input/target latent or pixel preprocessing; and
• behavior when the condition is missing, malformed, masked, or out of distribution.

Never silently drop an image condition and return an unconditional result as if the restoration succeeded.

Build honest degradation evidence¶

Synthetic degradation can create controlled paired examples, but it models only the transforms that were declared. Keep synthetic and observed real-world degradations distinguishable. Do not claim that noise, blur, compression, or downscaling parameters reproduce an unknown camera or editing history without evidence.

Split training, tuning, and holdout data by original asset, capture session, subject/product, and derivative chain. Adjacent crops or multiple degraded versions of the same original belong in the same split.

Evaluate recovery separately from plausibility¶

Review text, marks, product geometry, faces, protected regions, and material detail separately from aggregate metrics. A high visual-quality score can coexist with a factually altered output.

Release gate¶

Promote a paired-restoration model only with reproducible pairs, a model-compatible conditioning path, source-disjoint holdouts, source/output provenance, and a policy for uncertain or changed factual detail. If the model cannot distinguish recovery from invention for the intended use, release it only as a reviewed enhancement proposal or keep it out of that workflow.

Related pages¶

• image restoration survey
• synthetic dataset pipeline
• upscaler evaluation