# Colorize old photos: 11.3% perceptual color fidelity vs historian palette accuracy

Emily Patterson · November 11, 2026

> An 11.3% LPIPS reduction signals perceptual fidelity, not historian palette accuracy. Use diffusion drafts only with a ≥11.3% blind-set gain.

| Takeaway | Detail |
| --- | --- |
| 11.3% LPIPS reduction signals perceptual fidelity, not palette accuracy | Treat the reported 11.3% LPIPS reduction for text-conditioned diffusion over GAN colorization as a perceptual-distance advantage, not a historian-palette accuracy rate. |
| Use diffusion drafts only with a ≥11.3% blind-set LPIPS gain | Adopt diffusion as the draft generator only when it beats your GAN baseline by at least the reported 11.3% LPIPS reduction on a blind evaluation set. |
| Require region-by-region historian palette review | Every diffusion draft must pass region-by-region historian palette approval to verify archival color accuracy beyond perceptual fidelity. |
| Palette.fm free previews cap images at 500x500 pixels | Palette.fm’s free plan offers colorization at no cost but resizes images to a maximum of 500x500 pixels and adds a Palette logo watermark in the bottom right. |

As summarized in the takeaway table, the LPIPS result is a perceptual-distance measure, not a historian-palette accuracy rate.

The workflow gate and the region-by-region historian palette review are set out in the takeaways above.

![Sunlit wooden archive table with faded sepia photographs](https://static.mm-ais.com/article-images-ai/colorize-old-photos-11-3-perceptual-colo-ai-6b9d33fb.jpg)
Sunlit wooden archive table with faded sepia photographs

## How text-conditioned diffusion assigns hue

Text-conditioned diffusion treats colorization as a guided denoising process: the model starts from noise and iteratively refines an image while conditioning on both the grayscale input and a text signal. Any named implementation details—such as a specific GitHub project or a particular model architecture—should be checked against that model’s documentation rather than assumed from this overview. The resulting colors are plausible inferences from language and grayscale patterns, not measurements of the original reflectance, dyes, or pigments present when the photograph was made.

GAN baselines approach the same task through adversarial training, where a generator proposes colorized outputs and a discriminator pushes them toward the distribution of natural color images. As Sonja Grgic’s 2021 IEEE Access overview, *Grayscale Image Colorization Methods: Overview and Evaluation*, classifies, modern colorization spans a range from human-guided techniques to automated learning-based methods; diffusion and GAN systems both fall on the automated end but differ in how they form and constrain their predictions. Understanding this distinction matters because the mechanism that produces a convincing image is not the same as a mechanism that recovers historically specific color.

LPIPS (Learned Perceptual Image Patch Similarity) evaluates these outputs as a perceptual-distance score in a learned feature space. A lower LPIPS score indicates that a colorization is closer to a reference image as judged by that feature representation, which is useful for comparing draft quality across models. It does not certify that a uniform, vehicle, or textile matches a historian’s documented dye, pigment, or fabric sample; perceptual proximity and archival color accuracy are different targets.

When you use a text-conditioned diffusion model as a draft generator, treat every caption-driven hue as a semantic hypothesis. Check whether the prompt’s implied materials are consistent with visible grayscale cues—edges, shading, and texture—and flag regions where language could override weak visual evidence. This mechanism-level reading keeps the draft useful for visualization while making clear that any claim about historically correct palette requires evidence beyond what the model’s conditioning and chrominance regression can provide.

![Overcast museum corridor lined with framed monochrome photographs](https://static.mm-ais.com/article-images-ai/colorize-old-photos-11-3-perceptual-colo-ai-35354890.jpg)
Overcast museum corridor lined with framed monochrome photographs

## What the 11.3% number can prove

This section audits four numbers and limits on their own terms: the reported 11.3% LPIPS reduction, the reported 12.4% better-than-GAN comparison, the 20-year method review, and the 500×500 free-preview limit. The central distinction is that LPIPS is a perceptual-distance metric, so a reduction says the diffusion output was closer to the reference in that reported comparison, not that its hues match a historian's palette.

The 11.3% LPIPS reduction should be verified against the model paper and your own blind set before use. Do not treat any additional better-than-GAN percentage as an independent fact unless you can find it in the cited evaluation; GAN colorization remains common in published methods, so one reported advantage is not universal replication. A blind set can use the 11.3% value as a perceptual-distance gate only.

Source 2, Grgic's IEEE Access overview of grayscale colorization methods, frames colorization as a 20-year field whose main goal is to convince the viewer of authenticity. Its available abstract gives no historian-palette accuracy percentage, so any palette-accuracy claim needs separate evidence.

GAN colorization fails the archival test because its strength is visual plausibility, not verifiability. A uniform, vehicle, or building facade may appear convincing while reflecting a statistically likely hue rather than a documented one. When no source anchors the color choice, the output is an illustration, not an archival reconstruction.

Text-conditioned diffusion improves perceptual fidelity relative to GAN drafts, but that advantage does not establish historical accuracy. A result can score closer to human perceptual judgments and still assign colors that lack documentary support. For archival work, the critical check is whether each contested region can be defended with evidence, not whether the overall image feels more realistic.

A historian palette provides that evidentiary standard but rarely scales to full-image coverage on its own. The hybrid workflow resolves this trade-off: use diffusion only when it demonstrably outperforms your GAN baseline on a blind evaluation set, then submit the draft to a region-by-region historian palette audit. Treat every unaudited region as provisional, and promote colors to archival status only when a cited reference supports them.

## Costs, resolution, and audit time

I treat budgeting for archival colorization as an auditable workflow, not a per-click expense: use the free-preview resolution gate before making visual claims, obtain a written quote for every historian-review fee, and log per-region audit time. The 2021 IEEE Access overview and the Palette.fm free-plan documentation do not provide a historian-review fee, per-image price, or standard audit duration, so any defensible cost model must be built from documented checks rather than assumed figures.

The resolution gate comes first. According to Palette.fm, the free plan resizes images to a maximum of 500x500 pixels and places the Palette logo watermark in the bottom right. If the evidence crop you need to evaluate exceeds that 500x500 limit, the preview cannot support pixel-level claims about weave, skin texture, or insignia. Before judging those details, commission or produce an unscaled crop from the source scan so the review is based on original-resolution evidence rather than a downsized surrogate.

| Control | Required check | Record to keep |
| --- | --- | --- |
| Resolution gate | Confirm the evidence crop fits within the current free-plan limit and is free of watermark obstruction; if not, obtain an unscaled crop. | Source dimensions, preview dimensions, and the crop used for judgment. |
| Historian review | Request a written quote per image and per region before work begins. | Quoted scope, regions covered, and revision terms. |
| Audit time | Track active review time for each region during the palette audit. | Per-region time log tied to the reviewed image version. |

Because no grounded fee schedule exists, I require a written quote per image and per region and treat it as a project input, not an afterthought. I then compare that quoted review cost against the cost of one re-review caused by an unsourced hue: if a region must be reopened because its color choice lacks documentary support, the added review cycle is a direct, avoidable expense. Quoting at the region level makes that tradeoff explicit and prevents a single flat estimate from hiding the labor concentrated in uniforms, signage, or other evidence-dense areas.

Audit time should be logged with the same discipline. Record per-region audit time as work proceeds, linking each entry to the specific crop, draft version, and source note under review. Even without a benchmark duration, a consistent time log reveals where scrutiny concentrates, supports future scoping, and documents that every claimed color decision received a traceable review.

## What the 11.3% cannot establish

The most persistent misreading of the reported 11.3% LPIPS reduction is that it represents 11.3% more historically accurate color. It does not. LPIPS measures perceptual distance against a reference image, so the delta signals that a text-conditioned diffusion draft can look closer to a plausible color photograph than a GAN baseline. Historian-palette accuracy is a different claim entirely: it requires sourced evidence for dyes, pigments, fabrics, signage, or provenance that a perceptual metric cannot observe or verify.

This distinction matters because visual plausibility and archival fidelity often diverge. As Sonja Grgic notes in *IEEE Access* (2021), the main goal of colorization is to convince the viewer of the authenticity of the result, but viewer conviction is not the same as documented historical fact. A draft can satisfy a perceptual-distance threshold while assigning a period-inappropriate uniform shade, misreading a faded material, or inventing color where the source record is silent. Treating the LPIPS gain as an accuracy rate collapses that evidentiary gap.

| Condition | How the rule applies | Reason |
| --- | --- | --- |
| No period reference exists | Rule breaks | No sourced dye, pigment, or provenance record is available to confirm or reject the draft palette. |
| Damaged print obscures texture | Rule breaks | Loss of material detail prevents reliable identification of fabrics, finishes, or surfaces that govern color decisions. |
| Free preview is below the evidence crop | Rule breaks | A resized preview, such as Palette.fm’s 500×500-pixel free output, may hide insignia, trim, or labels needed for verification. |
| Caption names a generic scene, not a specific object | Rule breaks | Generic guidance cannot anchor color claims to a documented item, regiment, brand, or interior. |
| Known uniform, labeled signage, or documented interior | Rule still wins | Independent references allow a region-by-region check to confirm or correct the perceptually strong draft. |

In practice, I treat a diffusion output that clears the blind-set LPIPS bar as a perceptual candidate, not a finished historical record. When any break condition applies, the correct action is to label the result as interpretive or withhold color claims for the affected regions. When the win condition applies, the draft earns its place only after each verifiable region is matched against documented sources.

## Worked audit: 3 photos, 2 checks

In 2026, I run every colorization candidate through a three-photo audit with named examples: a portrait with a known uniform, an exterior with sky and fabric, and a damaged print. For each candidate, compare the diffusion draft with the GAN baseline, check whether a Palette.fm free preview is capped at 500×500 pixels, and review five named regions. Record the comparison, the evidenced regions, and the audit outcome with the image so later reviewers can see which rule governed the result.

Rules one and two separate evidence from absence. A dated uniform or object-specific reference makes the historian palette the first authority, with diffusion limited to non-evidenced background. With no period or object reference, the output must be labeled perceptual-only and cannot be published as historically accurate, even when LPIPS favors diffusion.

Rules three and four protect the metric boundary. If diffusion fails to clear the reported LPIPS reduction against the GAN baseline on a blind set, the diffusion advantage claim stops for that batch. If it clears the threshold but an evidenced region conflicts with the historian palette, the region is rejected or revised; the draft never overrides evidence.

Rule five fixes publication status. Archival use requires a region-by-region historian palette audit after the diffusion draft, regardless of the LPIPS result. Perceptual-only outputs remain useful for exploration, but they do not become archival claims through a better score.

Record the blind-set comparison, the evidenced regions, and the audit outcome with the image so later reviewers can see which rule governed the result.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | Assemble a blind evaluation set of your archival photos and colorize each image using both your GAN baseline and a text-conditioned diffusion model. | Creates a controlled comparison to measure perceptual differences between the two approaches on the same source material. |
| 2 | Calculate LPIPS scores for both outputs and adopt diffusion as the draft generator only if it achieves at least the LPIPS reduction benchmark noted above over your GAN baseline. | Ensures the diffusion model provides a meaningful perceptual fidelity advantage before it is used in your workflow. |
| 3 | Generate initial colorization drafts with the selected diffusion model, treating every output as a perceptual draft rather than a verified historical record. | Reinforces that lower perceptual distance does not guarantee accurate historical color palettes. |
| 4 | If using Palette.fm’s free plan for previews, account for its reduced image dimensions and reserve full-resolution colorization for your final, reviewed drafts. | Prevents resolution limits in preview tools from compromising the quality of images prepared for publication. |
| 5 | Conduct a region-by-region historian palette review, matching each area of the draft against documented archival color sources. | Verifies color accuracy at a granular level, addressing details that perceptual metrics alone cannot validate. |
| 6 | Publish the colorized image only after every region has received historian palette approval. | Ensures the final image meets standards for historical accuracy, not just perceptual fidelity. |

## Frequently Asked Questions

**Under what condition should text-conditioned diffusion be adopted over a GAN for colorization drafts?**

Diffusion should be adopted as the draft generator only when it beats your GAN baseline by at least the reported 11.3% LPIPS reduction on a blind evaluation set.

**Does a lower LPIPS score guarantee that a colorized photo matches historical color palettes?**

No, an LPIPS reduction signals a perceptual-distance advantage, not a historian-palette accuracy rate.

**What review process is required to verify the archival color accuracy of a diffusion draft?**

Every diffusion draft must pass region-by-region historian palette approval to verify archival color accuracy beyond perceptual fidelity.

**What is the maximum image resolution allowed for Palette.fm’s free plan previews?**

Palette.fm’s free plan resizes images to a maximum of 500x500 pixels.

**Does Palette.fm add any branding to images colorized on its free plan?**

Yes, Palette.fm’s free plan adds a Palette logo watermark in the bottom right of the image.

**How does text-conditioned diffusion assign color to a grayscale image?**

Text-conditioned diffusion treats colorization as a guided denoising process that starts from noise and iteratively refines an image while conditioning on both the grayscale input and a text signal.

## Quick answers

| What does the reported LPIPS reduction for text-conditioned diffusion over GAN colorization actually signal? | It signals a perceptual-distance advantage, not a historian-palette accuracy rate. |
| --- | --- |
| When should diffusion be adopted as the draft generator? | Only when it beats the GAN baseline by at least the reported LPIPS reduction on a blind evaluation set. |
| What must every diffusion draft pass before it is accepted? | Region-by-region historian palette approval to verify archival color accuracy beyond perceptual fidelity. |
| What are the limits of Palette.fm's free plan? | It colorizes at no cost but resizes images to a maximum of 500x500 pixels and adds a Palette logo watermark in the bottom right. |
| How does text-conditioned diffusion assign hue during colorization? | It treats colorization as a guided denoising process, starting from noise and iteratively refining the image while conditioning on both the grayscale input and a text signal. |

Also worth reading: **Whether diffusion-based colorization improves or harms historical authenticity in archival photographs: a side-by-side evaluation of color fidelity, luminance preservation, and expert-rated historical accuracy**: [Whether diffusion-based colorization improves or](https://colorizethis.io/blog/whether-diffusion-based-colorization-improves-or-harms-historical-authenticity-in-archival-photographs-a-side-by-side-evaluation-of-color-fidelity-luminance-preservation-and-expert-rated-historical-ac.php) · **Cyanotype colorization: DeOldify vs Stable Diffusion 2025, by provenance class**: [Cyanotype colorization: DeOldify vs Stable](https://colorizethis.io/blog/cyanotype-colorization-deoldify-vs-stable-diffusion-2025-by-provenance-class.php) · **A critical look at AI photo colorization**: [critical look at AI photo](https://colorizethis.io/blog/a_critical_look_at_ai_photo_colorization.php)

### Related reading

- [Colorize 1940s Kodachrome in 2026: 16-bit TIFF vs 8-bit JPEG skin-tone accuracy](https://colorizethis.io/blog/colorize-1940s-kodachrome-in-2026-16-bit-tiff-vs-8-bit-jpeg-skin-tone-accuracy.php)
- [Colorize Historical Photos: Verify Before You Commit to Refinement](https://colorizethis.io/blog/colorize-historical-photos-verify-before-you-commit-to-refinement.php)
- [Colorize Old Cotton Mill Photos: 15 Degree Lock vs Drift Wins 5-1](https://colorizethis.io/blog/colorize-old-cotton-mill-photos-15-degree-lock-vs-drift-wins-5-1.php)
- [How to Colorize and Restore Old Family Photos: The 2026 Definitive Guide](https://colorizethis.io/blog/how_to_colorize_and_restore_old_family_photos_the_2026_definitive_guide.php)
- [Diffusion Colorization: 92% Accuracy on Civil War Sepia Photos](https://colorizethis.io/blog/diffusion-colorization-92-accuracy-on-civil-war-sepia-photos.php)
- [Broken Camera? Still Colorize Your Old Black and White Photos](https://colorizethis.io/blog/broken_camera_still_colorize_your_old_black_and_white_photos.php)

### Latest

- [Blender Grease Pencil vs Toon Boom Harmony 2026: A Verify-Before-You-Commit...](https://colorizethis.io/blog/blender-grease-pencil-vs-toon-boom-harmony-2026-a-verify-before-you-commit-guide.php)
- [Colorize 1920s Studio Portraits: ControlNet vs DeOldify Skin Checks](https://colorizethis.io/blog/colorize-1920s-studio-portraits-controlnet-vs-deoldify-skin-checks.php)
- [MIT OpenCourseWare Migration: Verify Diffusion Model Fidelity for Daguerreotype...](https://colorizethis.io/blog/mit-opencourseware-migration-verify-diffusion-model-fidelity-for-daguerreotype-restoration.php)

Canonical: https://colorizethis.io/blog/colorize-old-photos-113-perceptual-color-fidelity-vs-historian-palette-accuracy.php
Markdown: https://colorizethis.io/blog/colorize-old-photos-113-perceptual-color-fidelity-vs-historian-palette-accuracy.php/index.md