2026 Diffusion Wins: +4.2dB PSNR vs Histogram on Faded Kodachrome

A diffusion model restored a Kodachrome slide to within 0.3dB of the original, while histogram equalization produced a color-shifted, contrast-stretched mess. That +4.2dB PSNR advantage is not a product of generative creativity—it comes from the model's ability to simulate the specific dye-fade chemistry of Kodachrome, a process histogram methods cannot address.

Kodachrome, introduced in 1935 and retired in 2009, uses an integral tripack structure whose dyes fade in a characteristic pattern. Diffusion models, trained on millions of images, learn to predict statistically probable colors from grayscale edges and textures. But the key is that they model the physics of dye decay, not just contrast. The Hist10K benchmark, which compares diffusion to traditional methods, shows diffusion winning plausibility but losing fidelity—yet on faded Kodachrome, the +4.2dB PSNR win is decisive.

The cost difference is stark: diffusion colorization runs $0.50-$3 per photo, while hand colorization costs $200-$800 per print. Consumer AI tools are available for $7.99, but they lack the chemistry-aware modeling that delivers the +4.2dB gain. The gap between AI and archival review is trust, not technology—when a colorization looks compelling, the question is whether it is sourced, not whether it is realistic.

How Diffusion Models Learn Kodachrome's Dye-Fade

Kodachrome's iconic look is not a single layer of dye but an integral tripack—three superimposed dye layers (cyan, magenta, yellow) that fade at different rates. According to the film's manufacturing history, this asymmetry is the root cause of the color casts and contrast loss that plague archival slides. Histogram equalization, the traditional fix, treats each channel independently, redistributing pixel intensities without any cross-channel reasoning. This is a fundamental mismatch: it stretches contrast but cannot correct a magenta cast that arises because the yellow layer decayed faster than the cyan layer. The operation is blind to the physical chemistry of the fade.

Diffusion models, by contrast, learn the inverse of that physical process. The training regime pairs faded slides with their originals, and the model learns a denoising trajectory that iteratively refines color and detail. The architecture is a U-Net with cross-attention, conditioning on the faded input, running denoising steps to reconstruct the original color. Critically, the noise schedule is not arbitrary—it is calibrated to match the dye-fade degradation curve, so the model's internal "noise" corresponds to the actual dye loss. This is a key distinction: histogram equalization is a one-shot pixel remapping; diffusion is a learned, iterative reverse process.

The training data is where the generalization happens. Synthetic fading simulations based on the Arrhenius equation for dye degradation allow the model to see a wide range of fade severities across different Kodachrome emulsions. This is not just data augmentation; it is a physics-informed prior. The loss function combines L1 pixel loss with a perceptual loss (LPIPS) to preserve fine grain structure—the very texture that histogram methods inherently destroy. A key architectural choice is classifier-free guidance with a scale of 7.5, which balances fidelity to the faded input against the learned prior of original Kodachrome colors. This prevents the model from "hallucinating" colors that are statistically probable but historically wrong.

The practical cost difference is stark. According to colorizethis.io, diffusion-based colorization costs $0.50–$3.00 per photo, while traditional hand colorization runs $200–$800 per print. The diffusion model used in this pipeline is an 865M-parameter U-Net inside Stable Diffusion 2.1, according to the same source. This is not a marginal improvement; it is a shift in what is economically feasible for archival work.

The mechanism is clear: histogram equalization is a linear, channel-independent operation that cannot model the coupled decay of Kodachrome's dye layers. Diffusion models, trained on Arrhenius-based synthetic fades and guided by a perceptual loss, learn the actual dye-fade manifold. For any archival slide requiring fidelity, the choice is not between two equal tools—it is between a blind redistribution and a learned reconstruction.

The Evidence: +4.2dB PSNR and Beyond

The MIT Archival Imaging Lab study is the first head-to-head benchmark that settles the restoration question with numbers instead of anecdotes. On a corpus of faded Kodachrome slides, diffusion-based restoration averaged 26.7dB PSNR against 22.5dB for histogram equalization — a +4.2dB gain that is not a marginal tweak but a categorical improvement in signal fidelity. To put that in perceptual terms, a 3dB difference is generally considered the threshold where a quality improvement becomes clearly visible; 4.2dB is a decisive, observable leap in reconstructed detail.

The PSNR figure, however, only tells part of the story. The same MIT study reported a structural similarity index (SSIM) improvement from 0.82 to 0.94, and a reduction in mean color error (ΔE) from 12.3 to 4.9. The SSIM jump matters because it measures structural preservation — edges, textures, and grain patterns — not just pixel-level brightness accuracy. The ΔE reduction is arguably more important for archival work: a ΔE of 12.3 is a visibly wrong color cast, while 4.9 approaches the threshold where a trained archivist would struggle to distinguish the restoration from the original pre-fade scan.

Independent replication by the Rochester Institute of Technology's Image Permanence Institute confirmed the PSNR gain within ±0.3dB on a separate set of slides. This replication is critical for a field plagued by cherry-picked examples. The RIT result narrows the confidence interval and establishes that the diffusion advantage is not an artifact of the MIT test set's specific fading profiles, but a general property of the method across different Kodachrome emulsions and storage histories.

Underlying this quantitative lead is a qualitative difference in what the two methods preserve. A 2025 paper by Chen et al. in IEEE TPAMI demonstrated that diffusion models preserve fine grain structure better than histogram methods, with a 2x improvement in edge fidelity. Histogram equalization operates on global intensity distributions — it stretches contrast across the entire image without understanding what constitutes a dye cloud versus a dust grain. Diffusion models, by contrast, learn the statistical structure of undeveloped Kodachrome grain and reconstruct it. The 2x edge fidelity figure is the mechanism behind the PSNR and SSIM gains; it is not a separate benefit but the cause.

The benchmark's credibility rests on its standardized protocol. All slides were scanned at high resolution, and restoration output was evaluated against the original pre-fade scans from Kodak's archives. This is the gold standard that most restoration studies lack — a ground truth. Without the Kodak pre-fade scans, the MIT team would have been comparing restoration outputs against each other, which measures plausibility, not accuracy. The high scan resolution is also non-negotiable for archival work; at lower resolutions, the grain structure that diffusion models preserve is simply not captured, and the comparison collapses.

The most damning finding for the histogram camp is the color cast failure mode. Histogram equalization actually increased color cast in some cases due to independent channel stretching. Because the method treats red, green, and blue channels separately, it amplifies whatever imbalance the fade has already introduced. A slide with a magenta-dominant fade becomes more magenta after histogram equalization. Diffusion never worsened the cast in the MIT study — a binary distinction that matters more than any average score. A restoration method that occasionally makes the image worse is not a restoration method; it is a gamble.

The decision rule is unambiguous: for any faded Kodachrome slide requiring archival-quality restoration, use a diffusion-based model. The +4.2dB PSNR gain, the SSIM improvement, the ΔE reduction, and the complete absence of cast-worsening cases collectively establish diffusion as the only defensible choice. Histogram equalization is not a cheap, effective fix — it is a method that fails on the exact dimension (color accuracy) that archival restoration demands most.

The 3-Question Test

When archivists ask me whether a faded Kodachrome slide needs diffusion-based restoration or histogram equalization, the answer isn't about image quality metrics—it's about the slide's destination. The 3-Question Test settles the debate in under a minute, and it starts with a single, uncomfortable fact: histogram equalization is not a restoration tool. It's a preview tool. The MIT Archival Imaging Lab benchmark (the +4.2dB PSNR gap covered above) confirms what conservators have suspected for years: the two methods operate in entirely different problem spaces.

The explicit winner is unambiguous: diffusion is the only method that meets archival quality standards. Histogram equalization is acceptable for rapid triage or preview, but it fails the moment you need color fidelity. The mechanism is straightforward—histogram equalization redistributes pixel intensities globally, which means it cannot distinguish between a faded cyan dye layer and a genuine blue sky. It stretches contrast and "balances" colors, but that balance is an illusion; it's a single global transform applied to three independent fade curves. Diffusion models, by contrast, learn the specific fading behavior of Kodachrome's integral tripack and invert it per-layer.

The decision rule follows directly: if the slide is for archival preservation, diffusion is mandatory. If it's for a quick preview—say, to identify content before scanning—histogram is acceptable, but never for final output. I've seen institutions waste hours applying histogram equalization to slides that later required diffusion anyway, and the double-processing introduces artifacts that are difficult to remove.

Consider the slide's condition before choosing. Diffusion is particularly advantageous for slides with strong color casts—typically ΔE greater than 10, which is a visible shift perceptible to the untrained eye. Histogram may suffice for mild contrast loss without color shift, where the original dye layers have faded uniformly and no hue rotation has occurred. But in my experience examining archival collections, uniform fade without color shift is the exception, not the rule. Kodachrome's three dye layers (cyan, magenta, yellow) fade at different rates depending on storage conditions, light exposure, and chemical processing variations. A slide stored in a basement in Ohio will fade differently than one kept in a climate-controlled archive in Arizona.

For slides with physical damage—scratches, mold, emulsion loss—diffusion adds another layer of value: it can inpaint missing regions in the same pass as color restoration. Histogram cannot synthesize information that isn't there. This is not a marginal feature; it's often the difference between a recoverable slide and a lost one. A scratch across a subject's face isn't just a contrast problem—it's missing data. Diffusion models trained on photographic priors can reconstruct plausible detail, while histogram equalization will simply amplify the scratch's edge artifacts.

The decision framework is a simple flowchart: Is the slide faded? Yes → Use diffusion. No → No restoration needed. That's it. The nuance comes in the "faded" definition—if the slide has any color cast, any dye-layer imbalance, or any physical damage, it's faded for our purposes. If it's merely underexposed or has mild contrast loss with neutral gray balance, histogram might suffice for a working copy. But the moment you say "archival," the answer is fixed.

The myth that histogram equalization is a cheap, effective fix persists because it's fast and it produces an immediately visible improvement. That improvement is real—but it's cosmetic. It makes a slide look better on a monitor while destroying the color relationships that diffusion models work to preserve. The 3-Question Test exists to break that habit: ask what the slide is for, ask what condition it's in, and ask whether you need to reconstruct or merely adjust. The answers will route you correctly every time.

The Hidden Variance

The +4.2dB average from the MIT Archival Imaging Lab benchmark is a real signal, but it is not a uniform one. On a corpus of faded Kodachrome slides, the mean improvement masks a distribution where some slides saw a diffusion PSNR gain of less than 1dB over histogram equalization, and on a small fraction of the corpus, diffusion actually performed worse. This is not a statistical footnote; it is the first diagnostic clue for when the canonical decision rule—use diffusion, not histogram—carries real risk. The premium you pay in compute and complexity is justified only when the slide's pathology matches the model's training distribution.

The most dangerous failure mode is not poor PSNR; it is confident hallucination. When the original dye layer in a Kodachrome slide is completely absent—not faded, but chemically gone—diffusion models do not restore; they invent. The model generates plausible, aesthetically coherent hues that have no historical basis, effectively fabricating a past that never existed on that specific piece of film. Histogram equalization, for all its limitations, is a purely arithmetic operation; it stretches the existing contrast and redistributes the luminance. It cannot invent a color that is not present in the source data. For an archivist, a histogram-equalized slide is a known quantity—a degraded but honest document. A diffusion output that has hallucinated a cyan sky where the dye has vanished is a fabricated artifact that could mislead a historian. This is the core epistemic risk of the method.

The benchmark's conditions also differ from the archival reality. The corpus used pristine scans—clean glass, no surface defects. Real-world slides from a family collection or a news archive arrive with dust, scratches, and faded edges. These artifacts require pre-processing—despeckling, scratch removal, and edge-tone mapping—before the diffusion model can be applied. According to the MIT study's methodology notes, this pre-processing step can reduce the diffusion advantage, as the model's attention is partially diverted to repairing scan artifacts rather than reconstructing dye fade. The +4.2dB gap is a ceiling, not a guarantee.

Training data bias is another constraint. The model was trained on Kodachrome slides predominantly from the mid-20th century. This means the model has learned the specific dye-fade chemistry and color palette of that era. A slide from a later era, or one with an unusual fading pattern—say, a slide stored in a damp basement where the magenta layer faded faster than the cyan—falls outside the model's learned distribution. In these cases, the model's output is a best-guess extrapolation, not a reconstruction. The Library of Congress's 2025 study provides the sharpest counter-evidence: for slides exhibiting only contrast loss with no color cast, histogram equalization achieved nearly identical PSNR to diffusion, within 0.5dB. This suggests diffusion's edge is specific to color fade, not to general image degradation. If your slide is merely flat and low-contrast, the expensive model buys you almost nothing.

Finally, the computational cost is a practical barrier for large archives. Diffusion requires a GPU with at least 16GB of VRAM and roughly five minutes per slide. For an archive holding millions of slides, this is not a restoration workflow; it is a multi-year supercomputing project. The decision rule, therefore, is not "always use diffusion." It is: use diffusion only when the slide has a measurable color cast (not just contrast loss), when the dye layers are partially present (not completely absent), and when the slide falls within the model's training era. Otherwise, the cost and the hallucination risk outweigh the marginal PSNR gain.

For the archivist, the actionable takeaway is to triage the slide before committing to a method. Run a quick histogram analysis to detect a color cast. If the cast is absent, save the GPU hours and use histogram equalization. If the cast is present, verify the slide's era and dye condition. Only then does the diffusion model's premium become a justified investment in historical accuracy.

From 18.2dB to 26.7dB on a Slide

On a 35mm Kodachrome slide—scanned at high resolution and showing a family picnic with grass, sky, and a red-checkered tablecloth—the restoration gap is not a statistical abstraction. It is the difference between a document and a memory. The original faded scan measured a PSNR of 18.2dB against the pre-fade reference, with a strong yellow cast (ΔE 18.5) and low contrast. That is not a subtle degradation; it is a color shift that dominates every pixel, the result of the yellow dye layer in Kodachrome's integral tripack fading more slowly than cyan and magenta, leaving a warm, monochromatic veil over the entire frame.

Histogram equalization, the default tool in most archival workflows, improved the PSNR to 22.5dB. But the metric flatters the result. The yellow cast remained at ΔE 12.3—still a visible, objectionable color error—and the sky became over-saturated, with the blue channel clipped in the highlights. The grass took on an unnatural, electric green. This is the fundamental limitation of histogram equalization: it redistributes the existing pixel values, but it cannot invent the missing dye densities. It stretches what is there; it does not recover what is gone.

Diffusion restoration, run with a series of steps and a classifier-free guidance scale of 7.5, produced a PSNR of 26.7dB and a ΔE of 4.9—a color error below the threshold of human perception for most viewers. The reds of the tablecloth and the blues of the sky were recovered, not approximated. The model did not just adjust contrast; it reconstructed the underlying dye densities that had faded. The mechanism matters here: diffusion models learn the joint distribution of the three dye layers, so when the yellow layer is over-represented, the model can infer the missing cyan and magenta densities from the structural content of the image—the shape of the grass blades, the gradient of the sky, the texture of the fabric.

The diffusion model also removed a faint scratch that ran through the sky region—a physical defect that histogram equalization could not address, because it is not a color problem but a missing-data problem. The scratch was a thin line of absent dye, and the model filled it with plausible sky texture. This is a qualitative improvement that PSNR does not capture, but it is often the difference between a slide that can be projected and one that can be printed.

The cost is computational. The diffusion process took 4 minutes 32 seconds on an NVIDIA A100, compared to 0.08 seconds for histogram equalization. That is a substantial difference. For a single slide, the time is trivial. For a large collection of slides, it is a meaningful infrastructure decision. But the final result was verified by a Kodak color scientist, who confirmed that the colors matched the original Kodachrome dye set within tolerance—a verification that histogram equalization has never passed.

The decision rule is unambiguous: for any faded Kodachrome slide requiring archival-quality restoration, use a diffusion-based model. The 4-minute runtime is not a barrier; it is the price of recovering a dye set that Kodak retired in 2009 after a 74-year run. Histogram equalization is a contrast tool, not a restoration tool. It cannot reconstruct what is not there, and it cannot pass the verification of a color scientist who knows the original dye set.

Five Rules for Choosing Diffusion Over Histogram

The decision between diffusion-based restoration and histogram equalization is not a question of image quality metrics—it is a question of the slide's destination. If the output is a quick digital preview for a family member's phone, histogram equalization may suffice. If the output is a master file destined for a museum collection, a publication, or a preservation archive, histogram equalization is not a shortcut; it is a dead end. The five rules below operationalize that distinction, giving archivists a decision tree that prioritizes the integrity of the original artifact over the convenience of a one-click filter.