Photo Colorization: Why PSNR 30 dB Is a Luminance Floor, Not a Color Guarantee

This guide distinguishes between luminance fidelity and color fidelity in AI photo colorization.

It explains when a PSNR of 30 dB can mask chroma errors and how to verify outputs using a Lab chromaticity check.

Why PSNR and expert ratings often diverge

PSNR is a mean squared error (MSE) metric based on pixel intensities. Since luma carries most of the signal energy in natural images, the PSNR score is overwhelmingly driven by luminance fidelity. By contrast, expert evaluations of historical colorizations place far more weight on chroma plausibility—especially skin tones, skies, and fabrics. As a result, PSNR and expert judgments can diverge even when the score reaches 30 dB. The key point: 30 dB indicates solid luminance fidelity, not a guarantee of accurate or believable color.

Because MSE averages squared differences across all pixels and channels, strong agreement in luminance can easily mask a hue shift. Even if that shift is modest in squared-error terms, it can be very noticeable to the human eye. A colorized image can preserve edges, textures, and tones while still introducing chroma that looks off, yet still score 30 dB simply because the luminance term dominates the average. Since PSNR is effectively calculated from luminance-dominant channels, its decibel value says very little about whether the chroma is plausible.

Expert review is a judgment of plausibility, not a pixel-distance test. Reviewers assess whether skin tones, skies, and fabrics look believable and historically appropriate—decisions that hinge almost entirely on chroma. For this reason, an image can hit 30 dB yet fail expert review over hue alone, even if its luminance reconstruction is nearly spot on.

he divergence can be screened with a Lab chromaticity check. A Photovoid-style analysis examines saturation distribution, chroma noise patterns, and edge color characteristics to separate real color from colorized black-and-white. The ~2 ΔE2000 figure used in this guide is a working convention, not a validated threshold from the sources; to check where divergence actually begins for your images, run the chromaticity check alongside PSNR and record the chroma error at which the two measures disagree.

This is the reader rule applied throughout the rest of the guide: the decision rules and worked example below show how the Photovoid-style Lab chromaticity check is paired with a PSNR ≥ 30 dB gate, and why the PSNR score alone counts as evidence of luminance fidelity only.

The check does not replace expert judgment; it makes the chroma failure measurable and repeatable. It turns a vague hue objection into a documented saturation or noise-pattern result, so a high PSNR can be qualified rather than mistaken for color authenticity.

Evidence: what the sources actually measure

Photovoid’s colorization detector is the named heuristic source in this evidence base. It returns a confidence score plus diagnostic details from four checks—saturation distribution, chroma noise patterns, edge color characteristics, and palette diversity—and it performs that analysis in Lab chromaticity space to separate a colorized black-and-white image from a true color photograph. Its history note also frames colorization as an interpretation rather than historical truth: manual colorization by artists tracing and hand-coloring photos dates to the 1860s, and digital colorization is done in tools such as Photoshop and GIMP. Any additional dating of digital or AI colorization eras should be verified against the source before being cited.

The capability claims come from the colorization vendors, not from validation studies. Imgupscaler.ai says its AI analyzes luminance, contrast, and object boundaries to apply context-aware color that respects historical accuracy. Palette.fm says it offers a new generation of AI colorization, with unlimited color previews, a free plan at no cost, and images resized to a maximum of 500. Those statements describe product behavior and access terms; they do not, by themselves, establish how a colorized result will score on a fidelity metric or on expert review.

What these sources actually measure is therefore narrower than a full fidelity verdict. Photovoid measures statistical chromaticity signatures that can indicate colorization; the vendor pages describe generation features and plan limits. Neither source supplies a PSNR threshold, an expert-rating protocol, or a causal link between a given chroma error and human judgment. Used together, they support a check—run the Lab chromaticity diagnostics before trusting a high-PSNR colorization—but they do not replace a controlled comparison of luminance and chroma outcomes.

In practice, the evidence base starts with a named set of checks: use Photovoid’s four diagnostics as the chromaticity check, treat vendor capability language as description rather than proof, and keep any PSNR or expert-rating claim tied to its own measurement method. That separation prevents a high luminance score from being mistaken for a color-fidelity guarantee.

Options compared: PSNR-only vs PSNR+chroma check

Option A is a PSNR-only gate at 30 dB. It is fast and reduces acceptance to a single number, but it is blind to hue error; it is acceptable only when the archive treats luminance fidelity as sufficient. It is useful for triage, not final acceptance.

Option B requires PSNR at or above 30 dB and a passing Photovoid-style Lab chromaticity check. It is slower because it has two stages, but it catches chroma failures that PSNR misses; for historical or archival use, it is the stronger gate. The extra stage is justified whenever hue error matters.

Option C relies on expert rating alone. It has the highest fidelity to human judgment, but it is not reproducible at scale and it is not a numeric gate.

The practical difference is where failure is caught. Option A can pass an image whose luminance is close to the reference while chroma is wrong; Option B stops that image at the second stage; Option C may catch it too, but only where a reviewer is available. That second stage is the difference between a luminance pass and a colorization that can be trusted in an archive. For a mixed archive, run Option A on low-risk batches, then promote any image that must be trusted to Option B, and reserve Option C for disputed or high-value items.

Winner: Option B — PSNR ≥ 30 dB plus a passing Photovoid-style Lab chromaticity check — because it keeps a numeric floor while adding the chroma screen that historical and archival review needs. Use Option A only when luminance fidelity is accepted as sufficient; otherwise require Option B before trusting a colorization at PSNR ≥ 30 dB.

Costs and numbers that matter

Free-tier palette.fm caps the long edge at 500 px, and its free plan offers colorization at no cost while resizing images to that maximum. A 4000×5000 archival scan must therefore be reduced before colorization: 8× on the 4000-pixel edge and 10× on the 5000-pixel edge. A PSNR computed on that downscaled output measures the small image, and it does not transfer to the full-resolution deliverable. Record the measurement dimensions with the score, or re-measure after the final upscale.

A 4000×5000 16-bit TIFF is about 120 MB, while the same frame as an 8-bit JPEG is about 8 MB. That is a 15× storage multiplier, so retaining PSNR-verified 16-bit masters costs roughly fifteen times the JPEG path in bytes. If storage is constrained, decide whether the verification workflow needs the larger master or only a JPEG-sized proof; the comparison is a byte-count ratio, not a price.

Photovoid’s detector returns a confidence score plus diagnostic details from four checks: saturation distribution, chroma noise patterns, edge color characteristics, and palette diversity. Because it is a heuristic, treat any single score as a screening signal rather than a calibrated probability. Corroborate it with the diagnostic details and, where possible, with the Lab chromaticity check before making a keep-or-reject decision; a score without its diagnostics is under-specified.

Practical accounting: state PSNR with its pixel dimensions, keep the 16-bit master only when the ~15× storage cost is justified by the verification need, and treat a Photovoid score as one cue among several. Use the same resolution for every comparison in a batch, and remember that these figures define the cost of measuring and storing evidence rather than settling color fidelity.

What the evidence does NOT establish

The 30 dB plus Lab chromaticity rule is a screening shortcut, not a validated acceptance test. It can fail in three specific ways, and each failure should move a result from "trusted" to "needs review."

First, no grounded source in this evidence base publishes a PSNR-to-expert-rating mapping, so 30 dB is a convention rather than a validated threshold. A colorization can clear 30 dB while still receiving a low expert rating, or fall below it while remaining acceptable, because the threshold has not been tied to rated outcomes. Check: before using 30 dB as a gate, require a documented mapping on the same image class and rating protocol; otherwise record PSNR as diagnostic only.

Second, Photovoid's heuristics are tuned to distinguish colorized black-and-white images from true color film, not to score colorization quality. That means the detector can answer a provenance question without answering a fidelity question. Edge case: a historically plausible colorization may still carry the colorization signature, while an atypical but authentic color photograph may be flagged. Rule: use the Lab chromaticity check for authenticity triage, then require separate chroma review before any quality claim.

Third, free-tier outputs resized to a maximum of 500 px, as palette.fm's free plan does, cannot be used to certify archival-resolution PSNR. Resizing and preview compression change the error surface, so a preview score does not transfer to the delivered file. Check: measure PSNR only on the final-resolution output, and label any 500 px result as preview-only until the full-resolution file is tested.

These limits also mean the ~2 ΔE2000 chroma-error trigger should be treated as a convention unless it is paired with rated data. If a workflow needs a hard gate, combine the full-resolution PSNR, the Photovoid-style chromaticity result, and an explicit expert-rating criterion rather than relying on any single number.

When any of these three conditions applies, keep the result in review status and do not cite PSNR or the chromaticity result as certification.

Gating a 4000×5000 scan

Begin with a 4000×5000 16-bit TIFF scan of a confirmed black-and-white original, roughly 120 MB, and treat the deliverable as an authenticity-labeled colorization rather than a preview. Keep the scan, the held-out reference, and the output in one tracked job folder, and store the reference outside the production folder so it cannot be overwritten by a later pass.

Checkpoint 1 is a full-resolution PSNR comparison against that held-out reference colorization. Compute it on the native 4000×5000 output, record the value with the file name and date, and stop if it is below 30 dB. Do not substitute a JPEG proxy or resized preview for this check; a failed value means the colorization does not clear the luminance-fidelity floor for this workflow, so do not send it to chroma review or label it as passing.

Checkpoint 2 is the Photovoid-style Lab chromaticity check on the same full-resolution output. Record the confidence score and all four diagnostic fields, then apply the reader rule: if chroma noise patterns or saturation distribution match the colorization signature, the output fails the authenticity label even when PSNR is at or above 30 dB.

If Checkpoint 1 fails, record the failed value and stop. If Checkpoint 2 fails, keep the PSNR result but withhold the authenticity label and route the file for chroma correction or manual review.

Only an output that passes both checkpoints may be marked as a screened, dual-gate-passed colorization; because the chromaticity check is a heuristic and no source validates the 30 dB threshold against expert ratings, the label records that both gates passed rather than certifying historical color authenticity. Archive the recorded PSNR, confidence score, diagnostic fields, and exact file versions with the TIFF so a later reviewer can reproduce the gate without reconstructing the workflow.

Use the same held-out reference and the same full-resolution settings for every rerun, and note any change to either in the job record before comparing new results.

Worked Example: Run the Numbers

Worked example (illustration only): in 2026, an archivist reviews a colorized historical portrait delivered as a 500-pixel-wide image, the free-plan maximum stated by palette.fm. The file’s reported quality is PSNR = 30 dB. The archivist also runs a Photovoid-style Lab chromaticity check and records mean chroma error = 2.4 ΔE2000 plus a colorization signature in the saturation/noise pattern. The decision question is whether the 30 dB score alone is enough to trust the color.

Calculation 1, threshold gap: 2.4 ΔE2000 − 2.0 ΔE2000 = 0.4 ΔE2000. Calculation 2, relative excess: 0.4 / 2.0 = 0.20, or 20%. Calculation 3, ratio to threshold: 2.4 / 2.0 = 1.2. These three results describe the same miss: the measured chroma error is above the ~2 ΔE2000 level at which expert rating is expected to diverge from PSNR.

Decision for this illustration: because the chroma error is 0.4 ΔE2000 above the ~2 ΔE2000 level and the signature is present, the 30 dB result does not settle the review in favor of trusting the color. The chroma-aware check is the controlling result for this example.

Break-even trigger: the added check stops changing the decision when the measured chroma error is at or below ~2 ΔE2000 and no colorization signature is reported. In that case, PSNR ≥ 30 dB and the chroma check agree, so the reviewer can treat the 30 dB result as sufficient for this 500-pixel-wide example. If either condition fails, as it does here with 2.4 ΔE2000 plus a signature, the chroma check remains the deciding test.

Decision rules

As established above, PSNR ≥ 30 dB is a luminance-fidelity floor only; the mandatory chromaticity check, the label wording for each outcome, and the batch-consistency requirement are set out in the decision-rules table and the gating walkthrough above.

If PSNR is at least 30 dB and the chromaticity check is clean, the image may ship as a screened colorization, not as a certified authentic color photograph. A clean result means neither chroma noise patterns nor saturation distribution match the colorization signature, but the check is a heuristic and the 30 dB-plus-chroma rule is a screening shortcut, not a validated acceptance test; record the PSNR value, check status, reviewer, and final label in the release record so the decision can be audited later.

If PSNR is at least 30 dB but the chromaticity check flags a colorization signature, do not make an authenticity claim. Send the image for expert review and label it as unverified or colorized-pending-review until a reviewer clears or rejects it; do not let a high PSNR value override the flagged chroma result or shorten the review.

If PSNR is below 30 dB, do not ship the image under any chromaticity result. Re-run colorization at full resolution, then repeat both gates from the start; a clean chroma result cannot compensate for a failed luminance floor, and a flagged result only adds another reason to withhold release.

Apply the rules in order and keep one current label per image. If reprocessing changes either gate result, replace the old label and note the reason, so downstream users never see a stale authenticity status or two conflicting labels on the same file. When the two gates disagree, the chromaticity result governs the wording of the label, while the PSNR result governs whether the image may be released at all.

Frequently Asked Questions

What metric is PSNR based on?

PSNR is a mean squared error (MSE) metric based on pixel intensities.

Why is PSNR overwhelmingly driven by luminance fidelity?

Since luma carries most of the signal energy in natural images, the PSNR score is overwhelmingly driven by luminance fidelity.

Which color aspects do expert evaluations of historical colorizations weight more heavily?

Expert evaluations of historical colorizations place far more weight on chroma plausibility—especially skin tones, skies, and fabrics.

Can PSNR and expert judgments diverge at 30 dB?

PSNR and expert judgments can diverge even when the score reaches 30 dB.

What does a 30 dB PSNR indicate?

30 dB indicates solid luminance fidelity, not a guarantee of accurate or believable color.

Can a modest hue shift still be noticeable?

Even if that shift is modest in squared-error terms, it can be very noticeable to the human eye.

Quick answers

Also worth reading: Diffusion Colorization: Conditioning Tradeoffs Beyond PSNR/SSIM: Diffusion Colorization: Conditioning Tradeoffs Beyond · A critical look at AI photo colorization: critical look at AI photo · How machine learning brings historical black and white photos back to life: How machine learning brings historical