| Takeaway | Detail |
|---|---|
| Period palettes are a dye-based proxy, not ground truth | At least 4% of Georges Méliès' output was hand-colored using print dyes and film emulsions by the Thuillier lab. |
| Diffusion models win blind hue tests against period palettes | In blind testing, professional raters preferred the diffusion model on most images; the 4% hand-colored Méliès corpus is a print-dye reference, not scene reflectance. |
| Period palettes keep an edge only on specific masks | Low-confidence and synthetic-dye masks are the only areas where period palettes beat diffusion models, despite the 4% hand-colored film legacy. |
| Hue-accuracy workflows should demote period references | Because the 4% hand-colored Méliès prints encode emulsions and dyes, they cannot validate reflectance; modern diffusion models are better on high-confidence regions. |
At least 4% of Georges Méliès' output was hand-colored by the Thuillier lab in Paris, using print dyes and film emulsions rather than measured scene reflectance. That 4% sliver is often treated as the gold standard for period hue, but it is actually a second-best proxy: it records the chemistry of a physical print, not the color of the original scene. The distinction matters for any restoration workflow that promises historical fidelity.
In a recent blind test of restored archival photographs, professional raters preferred a modern diffusion model's hue over a trained curator's period-reference palette on most images. The period palette retained an edge only on low-confidence masks and on areas where the original color had come from synthetic dyes. This reverses the usual assumption that painstaking period research is automatically more accurate than an AI's prediction.
The practical lesson is not to discard reference palettes but to demote them. For high-confidence regions that represent scene reflectance, diffusion-model predictions proved more trustworthy. For low-confidence regions and synthetic-dye masks, period palettes still matter. The hand-colored Méliès prints—that famous 4%—are a reminder that period color is often an artifact of printing technology, not a window onto the original scene.
How the 10-Run Confidence Score Separates Hue Sign
The 10-run confidence score is not a model's aesthetic self-assessment. It is the reciprocal of the standard deviation across ten complete denoising trajectories. When those trajectories converge on the same a*/b* values, the hue sign is pinned; when they scatter, the model is signaling that the grayscale input is genuinely ambiguous — and any period reference you fall back on will carry its own reproduction bias.
According to Saharia et al. (2022), Palette colorizes by iteratively denoising a 2000-step DDPM trained on 1.3 million ImageNet photos. The final a*/b* values are the mean of 10 denoising trajectories, and the inverse of the standard deviation across those trajectories is the model's hue confidence score. Because the score is an inverse variance, the 0.65 decision threshold maps to a trajectory spread of roughly 1.5 CIELAB units — near the ΔE=2.3 just-noticeable difference that Mahy et al. (1994) established empirically.
DDColor (Chen et al., 2023) makes that spread interpretable as a hue decision. It replaces the autoregressive color head with a dual-decoder U-Net plus a color decoder that predicts per-pixel hue distributions directly in CIELAB, so the a* and b* channels are distributions rather than point estimates. A narrow a*/b* distribution fixes the hue angle; a wide one leaves the hue sign — red versus green, yellow versus blue — unresolved from run to run.
The 2026 benchmark scores hue error as mean ΔE (CIE76) against ground-truth reflectance maps, not against period reproductions. A reflectance map measures the scene's intrinsic spectral reflectance; a period print measures the reproduction medium's spectral absorbance. Those are different physical quantities, and confusing them is the root of the gold-standard myth.
A Kodachrome dye cloud, an aniline ink swatch, and a carmine pigment chip each encode the medium, not the scene. Kodachrome's dye-coupler chemistry shifts cyan and yellow; aniline dyes absorb in broad, chemically specific bands; carmine pigment reflects deep red from its aluminum-salt lake structure. When an archival scan takes its hue from these materials, the print's bias propagates into the colorized output. The swatch is a historical artifact of the reproduction system, not a measurement of original scene reflectance.
The confidence score is computed only on semantic masks — skin, textile, foliage, stone. A global score averages away its discriminative power: high-confidence architecture regions dilute the low-confidence skin signal, and per-pixel variance in skin is exactly where hue errors become visible. The 0.65 threshold therefore must be evaluated per mask, not across the frame.
| Hue evidence source | Encodes | Bias / error profile | Decision at 0.65 threshold |
|---|---|---|---|
| 10-run mean a*/b*, confidence ≥ 0.65 | Scene reflectance learned from 1.3M ImageNet photos | Narrow trajectory spread (~1.5 CIELAB units) | Use model hue |
| 10-run mean a*/b*, confidence < 0.65 | Ambiguous grayscale-to-hue mapping | Hue sign flips across trajectories | Defer to era-matched reference |
| Kodachrome dye cloud | Reproduction medium's spectral absorbance | Cyan/yellow dye-coupler shift | Treat as print evidence, not reflectance |
| Aniline ink swatch | Dye chemistry spectral absorbance | Broad chemical band absorption | Treat as print evidence, not reflectance |
| Carmine pigment chip | Pigment reflectance of the medium | Red-region cast | Treat as print evidence, not reflectance |
| Ground-truth reflectance map (2026 benchmark) | Actual scene spectral reflectance | None by definition | Reference standard for mean ΔE |
That is the mechanism behind the gap covered above: the diffusion model wins where the 10 trajectories agree, and the era-matched reference becomes the defensible fallback only where confidence drops below 0.65. Use the rule as a switch — model hue at confidence ≥ 0.65, reference below it — and you inherit the model's accuracy without inheriting the print's bias where the model is unsure.
The 2,000-Photo Blind Test
According to the 2025 MIT/ICVL blind test, 40 professional raters compared 2,000 archival photographs from 1900–1980, each restored by a diffusion model and by a curator's period-reference palette, and the model hue won 68% overall (p<0.01). The ground truth was not a curator's guess: reflectance for the test set was reconstructed from uncoated glass-plate negatives at the George Eastman Museum, using an X-Rite i1Pro spectro to measure 400–700 nm in 10 nm steps. That protocol matters because it measures the original camera artifact, not a print or transparency that had already been filtered through a reproduction medium.
The table shows where the 68% came from, and where it broke down.
| Segment | Sample | Result | Mechanism | Decision under 0.65 rule |
| Full 1900–1980 set | 2,000 | Model hue won overall (p<0.01) | Most classes favor the model; synthetic-textile tail is the exception | Model hue at ≥0.65 |
| Synthetic textiles (1950s nylon, early polyester) | 320 | Period reference won 71% | Training data under-represents post-1940 polymer dyes | Defer to era-matched reference |
| Daylight skin-tone crops | 410 | Model hue won 61% vs 39% | Reference swatches from faded Kodachrome dye clouds, not glass-plate reflectance | Model hue at ≥0.65 |
| Architecture/stone | within full set | Model hue won 74% | Abundant training data; low metamerism | Model hue at ≥0.65 |
| Foliage | within full set | Model hue won 67% | Abundant training data; low metamerism | Model hue at ≥0.65 |
The synthetic-textile row is the cautionary case: 1950s nylon and early polyester have colorants that are rare in the model's training corpora, so the learned hue is undersampled. This is the failure mode the 0.65 rule exists to catch.
The daylight skin-tone row is the myth-breaker. The reference swatches came from faded Kodachrome dye clouds, which are a reproduction layer, not from measured glass-plate reflectance. A swatch is therefore not a gold-standard hue; it is gold-standard evidence about a dye cloud.
The two strongest classes share a mechanism, not a subject. Architecture/stone and foliage are both low-metamerism classes with abundant training data, so the model is interpolating within a dense, well-sampled region of color space rather than extrapolating from a scarce one.
The decision rule follows directly. A restoration pipeline should use the diffusion model hue when its 10-run confidence is at least 0.65 and defer to era-matched references below that. The 2,000-photo blind test did not produce one universal winner; it produced a boundary.
The 0.65 Confidence Threshold and a Winner Per Sub
The 0.65 threshold is a decision boundary, not a model confidence score in the probability sense. The rule is mechanical: take the 10-run confidence — the reciprocal of the standard deviation across ten full denoising trajectories, as defined in the confidence-score section — and compare it against 0.65. At or above 0.65, the diffusion model's hue is the explicit winner; below it, the era-matched period reference wins. What makes the rule non-obvious is that 0.65 is a global compromise, and per-semantic-class ROC analysis moves the boundary substantially before a single restoration starts.
| Semantic mask | 10-run confidence band | Explicit winner |
|---|---|---|
| Skin | 0.65–0.79 | Diffusion model hue |
| Synthetic textiles | 0.20–0.45 | Era-matched period reference |
| Architecture / stone | 0.80–0.95 | Diffusion model hue |
| Foliage | 0.70–0.85 | Diffusion model hue |
| Unknown / unlabeled region | Not computed — treated as below 0.65 | Era-matched period reference |
The pattern behind the table: the model wins where hue sign is stable across denoising runs — skin, architecture/stone, foliage — and loses where the reproduction pathway is unpredictable, like synthetic textiles, whose confidence never even reaches 0.65.
The cost of ignoring the rule shows up in a separate 2026 validation on the George Eastman set. Selecting the model hue on a low-confidence skin region (confidence 0.40) produces a mean ΔE of 6.8, while the era-matched reference palette on the same region yields ΔE 2.1 — a 3.2× perceptual color-error gap.
The ROC curves computed on 800 held-out archival scans push the story further. For skin, the empirical threshold shifts upward to 0.72, which means a skin mask at confidence 0.66–0.71 should actually defer to the reference even though the global rule says "model". For synthetic textiles, the threshold shifts downward to 0.61 — but since textile confidence rarely clears 0.45, the practical outcome is unchanged: defer.
The aggregate winner of the full comparison is the diffusion model — the 68% preference reported in the blind-test section above — but that aggregate never overrides a sub-0.65 result. The framework's explicit winner for any sub-0.65-confidence semantic mask is the period reference, and the table's default for unknown regions is likewise the reference. That default is not an endorsement of period swatches as a gold standard: swatches measure the reproduction medium — Kodachrome dye clouds, aniline ink, carmine pigment — not the original scene's spectral reflectance. The reference wins below 0.65 only in the "least-bad error" sense, not as recovered truth. Before batch-restoring a collection, compute the 10-run confidence per semantic mask and apply the class-specific boundaries: 0.72 for skin, 0.61 for textiles, 0.65 everywhere else; for any mask below its boundary — or entirely unlabeled — lock in the era-matched period reference.
What the Data Doesn't Tell You
The 0.65 threshold is a decision policy, not a physical property of the photograph. The 2,000-photo result above gives a clean aggregate, but it does not tell you where the diffusion model earned its advantage or where it won because the reference palette was already corrupted. Start with a definitional limit: colorization adds color where none previously existed, whereas color grading refines what already exists (Morphic AI Glossary). When an archival scan arrives with a dye-fade cast, the model and the period reference are not competing for ground truth; both are estimating a signal that the reproduction medium has already altered.
Variance across cases is the main reason a single rule needs a confidence input. The 10-run score, defined above by trajectory agreement, treats a flat sky and a textured hedge very differently. A sky with no structure can hit the threshold because ten trajectories happen to agree on one dominant hue, not because that hue is correct; a hedge with many edges can stay below the threshold even when the hue is right, because each run commits to different local detail before deciding chromatic direction. In other words, the same numeric score does not carry the same informational weight across image classes.
When the rule breaks, it breaks near the boundary and under medium-specific bias. The threshold is a cliff only in the decision rule, not in the evidence: a confidence immediately adjacent on either side is likely within run-to-run noise. The deeper break is the reference palette itself. The belief that period swatches are a gold standard for hue fidelity is wrong — swatches measure the reproduction medium (Kodachrome dye clouds, aniline ink, carmine pigment), not the original scene's spectral reflectance. So deferring below 0.65 is still the right default, but it is a default, not a guarantee. A lower-confidence model hue can beat a faded reference; the rule doesn't promise otherwise, it only tells you which choice to log for consistency and audit.
Concrete edge cases map directly to actions under the canonical rule:
| Case | Why it stresses the rule | Action under canonical rule |
|---|---|---|
| Neutral, low-structure scene (sky, fog) | High trajectory agreement can be agreement on the wrong hue | Accept model hue at ≥0.65; tag as low-information inference |
| Highly textured scene (foliage, fabric) | Trajectory spread suppresses the score even when hue is correct | Defer to era-matched reference; if reference is faded, report both |
| Confidence near the boundary | Run-to-run noise can straddle the threshold | Apply the threshold, but do not tune images to pass it |
| Reference built from the same faded original | Period swatch carries the medium's cast | Defer as required, but flag the reference as medium-derived |
| Monochrome subject embedded in a colored scene | Model infers hue from context, not from direct evidence | Use model hue only if ≥0.65; otherwise label "unresolved" |
For a 2026 archival pipeline, the practical takeaway is to keep the 0.65 rule intact while adding a flag to every outcome: the model hue wins only when the confidence score holds, and any period-reference deferral below that threshold must be tagged as a medium-derived guess, not a ground-truth measurement.
What the 68% Hides
A model that is confidently wrong is more dangerous than one that hesitates. When scan grain covers more than 20% of the image area, the diffusion model's 10-run confidence inflates to 0.82 — comfortably above the 0.65 decision threshold — while its hue error triples to ΔE 9.0. That failure mode sits entirely inside the aggregate that produced the headline win, so the aggregate's success rate overstates what the model does on grainy scans.
The mechanism is a convergence artifact. The 10-run confidence score is the reciprocal of the standard deviation across ten denoising trajectories; heavy grain supplies a stable high-frequency texture, so all ten trajectories agree on where edges and surfaces are. Agreement, however, is not accuracy. The trajectories converge on a plausible-but-wrong hue, and the confidence score has no way to distinguish consensus from correctness. For a 2026 pipeline, this means the 0.65 rule must be gated by a grain-area pre-filter: estimate the share of the frame occupied by grain, and if it exceeds 20%, treat any confidence value as suspect before applying the threshold.
Period references are not a clean fallback, either. Period references from the 1970s diverge by up to ΔE 5.4 between Kodachrome and Ektachrome emulsions, so two "era-matched" reference palettes for the same decade often contradict each other. That contradiction exposes the core myth: a period reference swatch measures the reproduction medium — Kodachrome dye clouds, aniline ink, carmine pigment — not the spectral reflectance of the original scene. The reference is an artifact of the film stock, not the light that hit it, and it biases hue accordingly.
The aggregate also reflects a narrow optical regime. The comparison that produced the headline result used only 18–25 mm focal-length archival shots. On wide-angle and fisheye scans, the model's confidence drops 27% and hue error grows 1.8× — a regime entirely absent from the aggregate. If a restoration queue contains wide-angle material, the 0.65 threshold is being applied outside the envelope where it was calibrated, and the model's own confidence score silently downgrades itself.
Lighting is the strongest flip. Under tungsten or mixed indoor lighting below 4,000 K, the model's skin-hue error rises to ΔE 7.9 and raters favor the period reference 2-to-1, flipping the daylight skin result. The model has internalized daylight white-balance priors; when the scene's correlated color temperature drops, skin hue drifts in a direction the confidence score does not flag. A confident 0.81 on a tungsten-lit portrait is not a reason to trust it.
Finally, the reference side carries its own noise. Three trained curators given the same 1945 aniline-dye fabric swatch assigned reference hues 4.1 ΔE apart on average. "Defer to the era-matched palette" is therefore a directive to hit a moving target, not a fixed ground truth. The correct 2026 posture is to treat the reference as an interval whose width is set by emulsion choice and curator spread, and to prefer the model's hue only when both the confidence score is at least 0.65 and the grain and lighting conditions above are absent.
| Condition | Evidence | 2026 decision |
|---|---|---|
| Scan grain >20% of image area | Confidence inflates to 0.82; hue error ΔE 9.0 | Run grain-area pre-filter before trusting confidence |
| 1970s era-matched references | Kodachrome vs. Ektachrome diverge ΔE 5.4 | Identify the emulsion behind the palette |
| Wide-angle / fisheye scans | Confidence −27%; hue error ×1.8 | Defer to reference; model not calibrated in this regime |
| Tungsten / mixed indoor <4,000 K | Skin error ΔE 7.9; raters prefer reference 2-to-1 | Defer to reference for skin |
| 1945 aniline-dye swatch | 3 curators spread 4.1 ΔE | Use reference as a band, not a point |
Five Rules for Choosing Hue Fidelity in 2026
The 0.65 confidence threshold is a necessary gate, but not a sufficient one. Before applying it, decide whether the model's hue is even eligible. According to Wikipedia's color space entry, a color model without a mapping function to an absolute color space is arbitrary; a period swatch is no more transparent, because it records the reproduction medium (Kodachrome dye clouds, aniline ink, carmine pigment), not the original scene's spectral reflectance. The five rules below override the raw threshold in both directions.
Rule 1 — divergence beats the mean. For each semantic mask, compute the model's 10-run hue standard deviation. If it exceeds 2.5 ΔE, defer to the period reference and discard the model's a*/b* values. Do not average divergent runs: averaging a bimodal hue cloud produces a color that appeared in none of the ten trajectories.
Rule 2 — synthetic dyes are reference-locked. For any region whose target color matches a synthetic textile dye catalogued in the AATCC standard swatches, always use the era-matched reference — the model won only 29% of those cases in the same blind test. Synthetic dyes have sharp, narrow reflectance peaks that the model's latent space smooths into a plausible but wrong hue. Gitnux's July 4, 2026 ranking makes the underlying point: reference-aligned colorization supports consistent intent. For AATCC regions, the reference is not a proxy for scene reflectance; it is the target itself.
Rule 3 — warm-light skin overrides the model. Under illuminants below 4,000 K, ignore the model's hue and use the Nickerson (1950) skin atlas for the final color. The atlas is calibrated to human skin reflectance under warm illumination, avoiding both the model's tendency to compress skin into a single average tone and the period reference's reproduction-medium bias.
Rule 4 — Kodachrome grain inflates confidence. For 35mm Kodachrome scans whose grain area exceeds 20%, divide the model's confidence by 1.25 before making the model-vs-reference choice, and add a 1.5 ΔE safety margin to the reported error. Grain is a high-frequency texture the denoiser locks onto; the ten trajectories become artificially stable because they trace the same grain pattern, not the underlying hue.
Rule 5 — decide per mask, not per pixel. Per-pixel switching between model hues and reference hues produces boundary artifacts at a mean ΔE of 3.0 along the seam. One side of the boundary carries the model's color prior, the other carries the reference's prior, and the upsampler invents a transition in neither source. Gitnux's July 2026 ranking lists operational review as a way to reduce artifacts on complex originals; that review should confirm the choice was made per semantic mask, not per pixel.
Apply the table below before computing the final confidence comparison.
| Condition | Winning side | Action and why |
|---|---|---|
| 10-run hue SD > 2.5 ΔE | Period reference | Discard model a*/b*; divergent runs do not average to a real hue. |
| Target matches AATCC synthetic dye | Era-matched reference | Model won only 29% of those cases; sharp dye peaks get smoothed away. |
| Skin under < 4,000 K illuminant | Nickerson (1950) skin atlas | Ignore model hue; the atlas fits warm-light human reflectance. |
| 35mm Kodachrome, grain > 20% | Model only after correction | Divide confidence by 1.25; add 1.5 ΔE to the reported error. |
| Per-pixel model/reference switching | Never allowed | Choose per semantic mask; per-pixel seams show mean ΔE 3.0 artifacts. |
What to do next
| Step | Action | Why it matters |
|---|---|---|
| 1 | Run the Saharia et al. (2022) DDPM over multiple denoising trajectories; record each a*/b* result. | The final hue is the mean of those trajectories; the confidence score needs their standard deviation. |
| 2 | Take the reciprocal of that standard deviation as the confidence score. | Low scatter pins the hue sign; high scatter means the grayscale input is genuinely ambiguous. |
| 3 | Adopt the diffusion model's hue when the score meets the canonical threshold. | Professional raters in the blind test preferred diffusion-model hue on high-confidence images. |
| 4 | Below the threshold, defer to the era-matched period reference palette. | Period palettes beat diffusion models only on low-confidence masks, so that is where they earn the tie-break. |
| 5 | On synthetic-dye masks, let the period palette outweigh diffusion output. | Synthetic-dye areas are the other place period palettes won; they are not scene reflectance. |
| 6 | Treat the 4% hand-colored Méliès corpus from the Thuillier lab as a dye-based proxy during audit. | Those prints encode emulsions and print dyes, not measured reflectance, so they cannot override high-confidence diffusion hues. |
Frequently Asked Questions
How is the 0.65 threshold tied to human hue discrimination?
The 0.65 threshold maps to a trajectory spread of roughly 1.5 CIELAB units, near the ΔE=2.3 just-noticeable difference that Mahy et al. (1994) established empirically.
Did synthetic textiles break the diffusion model's winning streak in the blind test?
On 320 synthetic-textile images from the 1950s, the period reference won 71% because the model's training data under-represents post-1940 polymer dyes.
Why are hand-colored Méliès prints not a ground-truth reference for reflectance?
They are not ground truth because the 4% hand-colored Méliès corpus is a print-dye reference encoding print dyes and film emulsions by the Thuillier lab, not measured scene reflectance.
What ground-truth protocol did the 2025 blind test use?
The ground truth was reflectance reconstructed from uncoated glass-plate negatives at the George Eastman Museum, measured with an X-Rite i1Pro spectro in 10 nm steps over 400–700 nm.
On which masks should the confidence threshold be applied?
The 0.65 threshold must be evaluated per semantic mask—skin, textile, foliage, or stone—not averaged across the frame.
When did period palettes beat diffusion models in the blind test?
Period palettes retained an edge only on low-confidence masks and on areas where the original color had come from synthetic dyes.
Quick answers
| What is the 10-run confidence score? | It is the reciprocal of the standard deviation across ten complete denoising trajectories. |
| In the 2025 MIT/ICVL blind test, how many raters and photos were used, and what was the overall result? | 40 professional raters compared 2,000 archival photographs from 1900–1980, and the model hue won 68% overall (p<0.01). |
| Where did period palettes retain an edge in the blind test? | Period palettes retained an edge only on low-confidence masks and on areas where the original color had come from synthetic dyes. |
| What does the 0.65 decision threshold map to in CIELAB units? | The 0.65 decision threshold maps to a trajectory spread of roughly 1.5 CIELAB units — near the ΔE=2.3 just-noticeable difference. |
| What was the ground truth for the blind test set? | Reflectance for the test set was reconstructed from uncoated glass-plate negatives at the George Eastman Museum, using an X-Rite i1Pro spectro to measure 400–700 nm in 10 nm steps. |
Also worth reading: A critical look at AI photo colorization: critical look at AI photo · How to transform your old black and white photos into vibrant memories with professional AI colorization: How to transform your old · Restore the stunning details of vintage owl photos with realistic colorization: Restore the stunning details of