title: Image Upscaler Evaluation: Fidelity-First Protocol description: "Choose an upscaler by measured fidelity on the actual source class, not benchmark labels or a universal default; preserve source/output provenance, evaluate artifacts and factual detail, and keep generative outputs out of factual training targets." category: tools tags: [upscaling, super-resolution, image-restoration, evaluation, data-provenance, fidelity]

Image Upscaler Evaluation: Fidelity-First Protocol¶

Upscaling creates a derived image. It can improve presentation or make a training input usable, but it cannot recover unobserved factual detail. Choose a model from evidence on the actual source class and intended use, rather than from a fixed leaderboard, a hardware estimate, or a claim that one tool is hallucination-free.

Separate task classes¶

Real-ESRGAN is an official project for practical image/video restoration trained with synthetic degradation. SwinIR documents restoration tasks including classical, lightweight, and real-world super-resolution, denoising, and JPEG artifact reduction. HAT is a super-resolution research model. These names identify candidates; they do not establish a universal ranking, safe default, license conclusion, or performance number for your hardware and images.

Evaluation dataset¶

Build the evaluation set before selecting a model:

• split by original asset/source, not by adjacent crops or derived copies;
• include the image classes the project will actually receive;
• keep paired references only where the known degradation is honest;
• reserve a source-disjoint holdout that no tuning decision sees; and
• record rights, original hash, crop/resize path, model revision, settings, and output hash for every derived image.

Do not use a model's own promotional gallery as a substitute for a held-out test set.

What to measure¶

For paired references, report metrics with their exact reference, crop, color space, alignment, and aggregation rule. PSNR, SSIM, or perceptual metrics can describe reconstruction error in that controlled setting; they do not prove factual preservation for unpaired images.

For all source classes, review:

1. repeated or erased texture, ringing, aliasing, and compression artifacts;
2. text, logos, product geometry, facial features, and other factual detail;
3. crop and aspect-ratio changes; and
4. peak memory, wall time, retry behavior, and failures on the intended runtime.

Use a blinded visual review where the decision matters. Record disagreement instead of resolving it by choosing the prettier result.

Training-data boundary¶

For factual, catalog, or supervised targets, preserve the original source as the authority. An upscaled derivative may be used only under a documented policy that distinguishes:

• preprocessing for display or model input;
• validated paired reconstruction experiments; and
• generative or uncertain detail that must not become ground truth.

If an output changes a mark, text, edge, or material detail that matters to a label, hold it for review or exclude it from that target. Saving as PNG does not repair an earlier model alteration, and deterministic processing is not a guarantee of factual recovery.

Runtime integration¶

Verify every concrete integration at the installed version: model file digest, loader, device/dtype, tile policy, alpha/bit-depth handling, input/output colorspace, and failure path. A ComfyUI node, a Python script, and a standalone binary are separate runtimes and require separate evidence.

Promote one candidate only after its source-class evaluation, provenance records, and operational measurement meet the intended release contract.

Related pages¶

• image restoration survey
• diffusion lora training
• flux klein 9b inference