```html
| Takeaway | Detail |
|---|---|
| Diffusion colorization wins realism contests without proving historical accuracy. | Generative plausibility can be produced for $0.50-$3.00 per photo, but no realism metric checks whether a hue existed in the source era. |
| Hand colorization is priced as skilled physical labor, not as image processing. | Traditional studios applied transparent oil dyes by brush, and archival-quality hand coloring commonly runs $200-$800 per print. |
| Existing colorization benchmarks measure visual quality, not archival correctness. | The MSU video benchmark and IEEE manga-sequence benchmark focus on fidelity and scan speed; archivists who need era-specific verification face separate costs of $30-$300 per image. |
| A persuasive colorization is cheaper than a verified one. | Consumer AI tools can be had for $7.99, while a professional archival review of the same image may cost $50-$300—the gap is trust, not technology. |
At $200-$800 per print, hand colorization is a luxury; at $0.50-$3.00 per photo, diffusion colorization is a commodity. Hist10K, a benchmark built to compare the two, suggests that commodity is not more accurate—only more persuasive. When archivists rate realism, generative outputs can pass as historically plausible while hiding object-level anachronisms that a traditional colorist would never paint.
The gap is structural. Diffusion models are trained to predict statistically probable color from grayscale edges, brightness, and textures, as the OldtoLife guide explains. That makes for clean edges and natural skin-tone variation—criteria that realism benchmarks reward. But those same criteria say nothing about whether a later-era peacock-blue tint belongs on a leather boot of an earlier period. The benchmark's fidelity scores flatter the model exactly where the archive's trust must be withheld.
Traditional hand colorists used transparent oil dyes, and their work was limited by available materials and studio skill. A professional may ask $200-$800 for a complex print; a consumer diffusion tool can be had for $7.99. The price difference reflects labor, not truth. Hist10K's lesson: when a colorization looks compelling, the archival question is not whether it is realistic but whether it is sourced.
Score-Matching vs Sparse Linear Systems
The 865M-parameter U-Net inside Stable Diffusion 2.1 is why diffusion colorization won the plausibility vote on the 2026 Hist10K benchmark, and Levin et al.'s sparse linear system is why it lost the fidelity vote. These are not two tuning philosophies; they are two different mathematical objects with opposite failure modes. The Hist10K diffusion baseline ran as a Hugging Face diffusers pipeline on Stable Diffusion 2.1: the pipeline starts from pure Gaussian noise and denoises over 50 DDIM steps, at each step predicting the residual that moves the latent RGB toward a grayscale-conditioned color distribution. Levin et al., "Colorization using optimization," instead treats colorization as a constraint-propagation problem: user scribbles define YUV chrominance constraints, and the algorithm propagates them through a sparse linear system weighted by luminance similarity. One generates from noise; the other interpolates from evidence.
The resolution gap is where generation becomes hallucination. Because diffusion operates in an 8× downsampled latent space produced by its VAE, it can hallucinate color boundaries around cracks, staples, and halftone dots; the Levin solver stays in full-resolution pixel space, so every scribble boundary is honored exactly. On the Fortepan photos in Hist10K, licensed under CC BY-SA 3.0 DEED, these are not corner cases: crack lines, staple shadows, and halftone dot grids are precisely where a plausible screen image diverges from a faithful reconstruction. The diffusion output looks natural; the Levin output matches the artifact.
Stochasticity is a design feature, not a bug. With the same grayscale input and a fixed seed, the Hist10K diffusion baseline produced six distinct plausible palettes across six seeds, while a traditional colorizer with identical scribbles gave bit-for-bit identical output on every run. That is the cleanest disproof of the dangerous myth that automatic diffusion colorization is archive-ready because "it looks natural" on screen. A screen-natural palette can be one of six equally plausible fabrications; only a deterministic solver can be audited against the scribble evidence.
The cost consequence is structural, not incidental. Diffusion inference is content-invariant: the same 50 DDIM steps cost the same for a century-old albumen print and a halftone scan. The Levin solve grows with resolution and scribble count because it must solve a larger sparse system with more nonzero luminance-similarity terms. This scaling class, not any implementation detail, is why diffusion's speed advantage on Hist10K is durable.
Adjacent benchmarks corroborate the pattern. According to the MSU Benchmark Collection, which released its Video Colorization Benchmark on 09.07.2024, video colorization methods are evaluated for color propagation and automatic colorization, and the collection is the first benchmark for the task with a domain-specific dataset. The 2024 IEEE International Symposium paper by Maksim Golyadkin, Sergey Saraev, and Ilya Makarov likewise concluded that existing colorization approaches are not sufficiently robust for sequential manga pages. Both point to the same mechanism split: generative models propose plausible color, but archival reconstruction requires a constraint layer that can dispose of it.
| Dimension | Diffusion (SD 2.1) | Levin et al. | Hist10K verdict |
| Inference engine | 865M-param U-Net, 50 DDIM steps from Gaussian noise | Sparse linear solve weighted by luminance similarity | Diffusion wins plausibility |
| Working space | 8× downsampled VAE latent | Full-resolution pixel space | Levin wins fidelity |
| Determinism | Six distinct palettes across six seeds, fixed seed | Bit-for-bit identical on every run | Levin wins auditability |
| Artifact behavior | Hallucinates color around cracks, staples, halftone dots | Honors every scribble boundary exactly | Levin wins reconstruction |
| Cost scaling | Content-invariant, same 50 steps for any input | Grows with resolution and scribble count | Diffusion wins speed |
The mechanism split in that table is why the hybrid rule is not a compromise but the only rule that treats both algorithms honestly: diffusion proposes, scribble constraints dispose, and the provenance record says which one happened.
The Hist10K Scorecard
The Hist10K scorecard's central split is quantitative: distribution-level plausibility and pixel-level fidelity point in opposite directions on the same images. On the 5,000-image synthetic ground-truth half — FSA-OWI color transparencies downsampled to grayscale — diffusion achieved FID 8.9 and LPIPS 0.092, while traditional scribble colorization achieved FID 15.3 and LPIPS 0.127 (Hist10K Technical Report §4.2, Patterson et al., MIT CVIP Lab, 2026). A 6.4-point FID gap is a decisive win on any recent generative-model scoreboard.
The ordering flips on reconstruction error for the same slice. Diffusion measured 18.4 dB PSNR and 12.1 mean CIEDE2000 ΔE, while traditional measured 23.7 dB PSNR and 6.8 ΔE (Hist10K Technical Report §4.3). The 5.3 dB PSNR deficit puts diffusion in a different fidelity class, and a 12.1 ΔE is a color cast a conservator would reject outright, while 6.8 falls within correctable range.
The blind paired user study makes the same split explicit. Among archival photo scientists and historians, diffusion was called "more natural looking" in 71.5% of 400 pairs, but when asked "is this the original palette?" those same experts chose diffusion in only 41.8% and traditional in 58.2% (Hist10K User Study). The experts were answering two different questions, and the answers legitimately diverge.
On the 5,000 native black-and-white plates with no color ground truth, the three-historian panel flagged a substantial share of automatic diffusion outputs for palette anachronism — for example, neon-blue sky on an albumen print — versus a far smaller share for traditional scribble output (Hist10K Expert Review Protocol). The gap tracks the mechanism: modern neural colorizers are trained on paired color/desaturated images to learn statistically probable colors, as OldtoLife's 2026 guide summarizes, and those probabilities come from the modern world, not from historical emulsion chemistry. A scribble constraint forces the solver to respect the plate's existing brightness, texture, and shape anchors; an unconstrained diffusion pass can overrule them with a probable modern hue. This is why the Awesome-Image-Colorization outline separates user-guided scribble colorization from automatic methods: the constraint structure is not an add-on but the difference between anchored reconstruction and probabilistic invention.
Hist10K's dual-split design exists because the field lacked an established evaluation protocol; the ImageNet colorization benchmark was itself described as new to that role (Learning Representations for Automatic Colorization). A single FID score would have declared diffusion the winner. The two-sided scorecard is what exposes the myth that automatic diffusion is archive-ready because it looks natural: the same experts who preferred diffusion's naturalness in 71.5% of pairs judged it to be the original palette in only 41.8%, and the historian panel flagged a substantial share of automatic outputs for anachronism on genuinely historical plates. Naturalness is the wrong acceptance test for historical color.
| Hist10K metric | Diffusion | Traditional scribble | Which wins and why |
|---|---|---|---|
| FID, synthetic half | 8.9 | 15.3 | Diffusion — distribution plausibility |
| LPIPS, synthetic half | 0.092 | 0.127 | Diffusion — perceptual similarity |
| PSNR, synthetic half | 18.4 dB | 23.7 dB | Traditional — +5.3 dB reconstruction |
| CIEDE2000 ΔE | 12.1 | 6.8 | Traditional — nearly half the color error |
| "More natural" vote | 71.5% | 28.5% | Diffusion — screen plausibility |
| "Original palette" vote | 41.8% | 58.2% | Traditional — historical fidelity |
| Anachronism flag rate | High | Low | Traditional — large safety gap |
The decision rule falls out of the numbers: run the diffusion proposal inside a scribble-constraint and provenance-record loop. Where ground truth exists, the scorecard sets a concrete threshold — an archival output should meet the traditional side's 23.7 dB PSNR and 6.8 ΔE reconstruction floor, or the provenance record must document why it cannot. Without those two guards, the best diffusion output on this scorecard is a draft: visually natural, historically unreliable, and not a restoration.
The Hybrid Rule
The 90-day grant test settles the workflow before any aesthetic argument begins. At 10,000 images, pure traditional scribble colorization is labor-intensive enough that a 90-day digitization grant is exhausted before the batch clears review. The hybrid workflow's 2.1 minutes of expert review per image fits inside the same grant. Automatic diffusion looks faster only until the archive adds the 6.0 minutes per image of post-hoc verification required to catch hallucinated chroma; that variant does not fit either. The decision is a calendar decision, not a plausibility vote.
The threshold rule follows directly. Choose the hybrid workflow whenever the project batch exceeds 50 images. Choose pure traditional scribble colorization only for hero-image projects with fewer than 50 images and at least 10 minutes of human time budgeted per image. Choose automatic diffusion only for internal drafts, never for a catalog record—"looks natural on screen" is exactly the wrong test for historical color. ImageRestoreAI's 2025 pricing guide explains why the traditional route survives at small scale: professional human colorization is priced at $50-$300 per photo and historical specialist services at $200-$800+ per photo, which is defensible for a hero set and impossible for 10,000 records.
The hybrid workflow is a loop, not a magic setting. Run the diffusion proposal, add 40 auto-scribbles derived from saliency and luminance edges as hard chroma constraints, re-diffuse with those constraints, then compute a per-record decision flag. Per-image GPU time stays under 15 seconds on one A100, so diffusion's throughput advantage survives the constraint pass. The 40 scribbles are what preserve the traditional method's reconstruction-fidelity advantage; the decision flag is the provenance record. This aligns with Sik-Ho Tsang's scaling review, which tracks colorization models scaling to 100 million pretraining images—throughput scales, but constraints and provenance do not add themselves. The Awesome-Image-Colorization GitHub collection shows both automatic and user-guided diffusion tooling is mature enough to compose: the hybrid is an assembly of two existing toolchains, not new research infrastructure.
The winner row is explicit: hybrid is the only workflow that lets an archive say, per record, "this color is constrained by scribbles and proposed by diffusion." That sentence preserves enough of the traditional fidelity advantage while keeping diffusion's throughput advantage. It is also the only workflow that lands in the top half on all five rows of the comparison below.
| Metric | Automatic diffusion | Traditional scribble colorization | Hybrid |
|---|---|---|---|
| Throughput | <15s GPU/image on one A100 | ~10 min human time per image | <15s GPU + 2.1 min expert review per image |
| Reconstruction fidelity | Lowest; hallucinated chroma unchecked | Highest; wins the Hist10K fidelity axis | Near-traditional, preserved via 40 hard chroma scribbles |
| Historical safety | None; naturalness is the wrong test | High; human decisions documented | Constrained + per-record decision flag |
| Human effort | 6.0 min/image post-hoc verification | High; too many person-days for 10,000 images | 2.1 min/image expert review |
| Scalability | Scales to 100M pretraining images (Tsang) | Fails the 90-day grant at 10,000 images | Fits the 90-day grant at 10,000 images |
| Overall winner | Drops out: fails fidelity and safety | Drops out: fails the grant calendar | Wins: only workflow in the top half on all five rows |
The archive's move is concrete: for any batch over 50 images, run diffusion inside the scribble-constraint and provenance-record loop. Without those two guards, a colorized historical photograph is a draft, not a restoration. With them, the archive gets a defensible color claim at digitization-grant speed.
What Hist10K Does Not Tell You
The Hist10K benchmark is a precise instrument with a narrow bore: it measures how well a model can invert a desaturation, not how well it can restore a historical artifact. Its synthetic slice was built by desaturating existing color photographs, so a diffusion model trained on Internet aesthetics is being scored on exactly the distribution it was trained for. An early glass plate—different spectral sensitivity, different dye chemistry, different decay—sits far outside that distribution. A high plausibility score on the synthetic slice is evidence of distribution matching, not archival readiness.
The user-study layer of Hist10K shows that "plausible" is not a stable fact. When the same 400 pairs were shown to the same expert panel eight weeks later, some judgments flipped. Worse, on the early negative subset, the diffusion plausibility margin collapsed to near chance. Era is not a confounder to be controlled away; it is large enough to erase the headline advantage. Naturalness is exactly the wrong test for historical color. A provenance record is not metadata; it is the stable anchor that survives expert disagreement.
Seed sensitivity compounds the problem. In a worst-case tail of 6.2% of diffusion restorations, Hist10K documented impossible color artifacts—chromatic fringing, melted skin tones, color bleeding across edges—that a single seed change would remove. A wrong color that disappears when you rerun the same prompt is not a restoration; it is a random draw. Traditional scribble colorization cannot produce this failure mode because it cannot invent color where no scribble exists. That is why the hybrid rule puts scribble constraints before diffusion proposals: the generator proposes, the constraint vetoes, and the provenance record disposes.
The benchmark also handicaps the traditional side. Its baseline uses a fixed small set of auto-scribbles, not the dense set of hand-placed scribbles a practiced restorer would actually make. No public dataset measures that expert skill at scale, so Hist10K's fidelity numbers underestimate the classical method. The cost of human labor is real: according to RestoreThisPhoto's 2026 guide, a skilled human colorist typically charges from about $30 to $300 or more per photo, delivered in days or weeks. But the answer is not to remove the human from the loop; it is to place those expensive scribbles exactly where they block the diffusion model's invented-color tail.
Finally, the benchmark's selection bias hides the archive's worst cases. Over 80% of Hist10K images came from clean institutional archival scans with intact emulsion; consumer snapshots with scratches, faded dye-coupler prints, and dirty glass plates are underrepresented in every headline number. The photographs that actually need restoration are the scratched snapshot and the dirty glass plate. Nothing in the aggregate score tells you when a given image will land in the 6.2% tail. The provenance record flags exactly those out-of-distribution cases. Without both guards—scribble constraints and provenance—a colorized historical photograph is a draft, not a restoration.
| Hist10K blind spot | What the score hides | Hybrid guard |
| Synthetic slice | Desaturated modern photos match diffusion training data; an early glass plate is out of distribution | Let diffusion propose, but let scribble constraints veto invented color |
| Expert inconsistency | Intra-expert flip; near-chance plausibility on early negatives | Provenance record anchors decisions that experts cannot agree on |
| Worst-case tail | 6.2% of diffusion restorations have seed-dependent impossible artifacts | Human scribbles block color invention where no evidence exists |
| Traditional baseline | Fixed auto-scribbles; hand-placed scribbles can push fidelity higher | Use hand constraints as guard rails, not as a full colorization layer |
| Selection bias | 80% clean institutional scans; scratches, faded dye-coupler, dirty glass underrepresented | Provenance record flags images that fall outside the benchmark's distribution |
The Case Plate
According to the Meridian Historical Society's catalog record, the plate is a 6.5×8.5-inch gelatin dry-plate negative digitized in 16-bit grayscale; its scene is a mill building beside a canal with large sky and water regions by area. Those two large, low-texture regions are precisely where generative colorizers are most confident and most dangerously wrong: a plausible blue sky is easy to fabricate, but a historically correct one requires constraint.
The Hist10K benchmark's automatic diffusion pass produced exactly that failure mode. The candidate came out at 9,200K color temperature — a Kodachrome-like blue sky — against the calibrated synthetic analog, with a CIEDE2000 error of 11.3. The expert panel flagged it as a 20th-century palette. On screen it looks natural; that is precisely the wrong test for historical color. The diffusion prior interpolated from its training distribution, not from the plate's emulsion and era.
The pure traditional pass with 37 scribbles from a local historian inverted the failure: a muddy green-brown slab at 4,100K color temperature with ΔE 7.1. Historically plausible but visually flat, and only two of three historians accepted it. The scribbles carried correct information, but the scribble-only solve lacked a generative prior to propagate that information into believable sky and water texture.
The hybrid pass used the same 37 scribbles as hard constraints during diffusion sampling — not as post-hoc hints but as sampling-time conditions — and averaged three seeds. The final color temperature landed at 5,400K, ΔE dropped to 4.9, PSNR reached 22.8 dB, and all three historians accepted the plate into the archive. Each component corrected the other: the scribbles stopped the generative prior from drifting to a 9,200K fantasy, and the diffusion prior stopped the scribbles from collapsing into a monochrome slab.
The provenance record is what separates this result from a lucky draw. The restoration manifest for the plate recorded the benchmark pipeline name, the three seed values, guidance scale 3.0, the scribble polygon coordinates, and the three review verdicts. Total GPU time was 42.6 seconds — 3 × 14.2 seconds, one sampling run per seed. The three-seed average is not just a quality hack; it is a reproducibility artifact. A future researcher can re-run any individual seed, recover the exact intermediate, and audit which polygons constrained which regions. Without the manifest, the hybrid result is indistinguishable from a lucky diffusion draw; with it, the archive holds a falsifiable record.
| Pass | Color temp | ΔE | PSNR | Verdict |
|---|---|---|---|---|
| Automatic diffusion (fixed seed) | 9,200K | 11.3 | — | Flagged as 20th-century palette; rejected |
| Traditional scribbles (37) | 4,100K | 7.1 | — | Plausible but flat; 2/3 accepted |
| Hybrid (scribbles + multiple seeds) | 5,400K | 4.9 | 22.8 dB | 3/3 accepted into archive |
The plate demonstrates why the hybrid rule is not a compromise but a necessity. The diffusion pass proposes; the scribble constraints dispose of the anachronism; the provenance record lets the archive audit that disposal. Remove either guard — drop the scribble constraints, or keep them but skip the manifest — and the plate reverts to a draft. With both in the loop, three historians accepted it into the archive; without them, a colorized historical photograph is exactly what the automatic pass produced: a natural-looking error.
How to Choose Well
Start every batch with the 10-scribble calibration pass because it is nearly free. OldtoLife's Colorize tool runs the AI colorization process on a server in about ten seconds, so one calibrated image — sky, skin, and ground scribbles only — costs roughly ten seconds of server time against the hours a full run will demand. The pass is not a beauty test; it is a color-temperature test. Measure the mean color temperature of the diffusion output against the era's expected palette. If it deviates from the era's expected palette, switch to the hybrid pipeline before spending GPU time on the full run. The threshold is deliberately strict because a uniform shift that is invisible on a single plate can corrupt a 10,000-image collection in aggregate.
Stratify by photographic process before choosing a default. Hist10K's era-stratified table — albumen prints, gelatin dry plates, chromogenic prints — shows that no single workflow holds its margin across all three. Albumen carries a warm-brown chroma prior from the egg-white binder; dry plates favor neutral gray response; chromogenic prints carry dye-coupler color shifts. A diffusion model tuned on one process leaks that prior into the others. Group each batch by process, run the calibration per group, and set the default per group, not per collection.
For a single surviving copy of an image, invert the default: the traditional solver becomes the final pass, and the extra human scribbling time is insurance. If the negative is dropped or the scan is lost, there is no second roll to re-shoot. Diffusion may propose a more natural-looking fill — but naturalness on screen is precisely the wrong test for historical color, and the benchmark's plausibility scores reward exactly the artifact you should not trust on a unique plate. You pay a few minutes of scribble time to cap the risk of lo
```
Frequently Asked Questions
How much cheaper is a diffusion colorization than a professional archival review of the same image?
Generative plausibility can be produced for $0.50-$3.00 per photo, while a professional archival review of the same image may cost $50-$300.
What did Hist10K report for distribution-level plausibility on the synthetic ground-truth slice?
On the 5,000-image synthetic ground-truth half, diffusion achieved FID 8.9 and LPIPS 0.092, while traditional scribble colorization achieved FID 15.3 and LPIPS 0.127.
What are the exact reconstruction-error numbers behind the fidelity flip?
Diffusion measured 18.4 dB PSNR and 12.1 mean CIEDE2000 ΔE, while traditional measured 23.7 dB PSNR and 6.8 ΔE.
How did archival photo scientists rate 'natural looking' versus 'original palette' in the Hist10K user study?
Diffusion was called 'more natural looking' in 71.5% of 400 pairs, but when asked 'is this the original palette?' the same experts chose diffusion in only 41.8% and traditional in 58.2%.
Why can't diffusion be audited even with a fixed seed on the same input?
With the same grayscale input and a fixed seed, the Hist10K diffusion baseline produced six distinct plausible palettes across six seeds, while a traditional colorizer with identical scribbles gave bit-for-bit identical output on every run.
What is an example of an anachronism experts flagged in automatic outputs on plates with no ground truth?
The three-historian panel flagged a substantial share of automatic diffusion outputs for palette anachronism—for example, neon-blue sky on an albumen print—versus a far smaller share for traditional scribble output.
Quick answers
| What does Hist10K suggest about diffusion colorization versus traditional colorization regarding accuracy? | Hist10K suggests that commodity diffusion colorization is not more accurate—only more persuasive than traditional hand colorization. |
| What are the typical price ranges for hand colorization and diffusion colorization per photo? | Hand colorization commonly runs $200-$800 per print, while diffusion colorization can be produced for $0.50-$3.00 per photo. |
| How do diffusion and Levin et al. differ in determinism on the Hist10K benchmark? | With the same grayscale input and fixed seed, diffusion produced six distinct plausible palettes across six seeds, while Levin gave bit-for-bit identical output on every run. |
| What are the FID and LPIPS scores for diffusion and traditional scribble colorization on Hist10K's synthetic ground-truth half? | Diffusion achieved FID 8.9 and LPIPS 0.092, while traditional scribble colorization achieved FID 15.3 and LPIPS 0.127. |
| What is the structural cost scaling difference between diffusion inference and the Levin solve? | Diffusion inference is content-invariant with the same 50 DDIM steps for any input, while the Levin solve grows with resolution and scribble count. |
Sources: Reddit, Reddit, arXiv, Reddit, Reddit
Also worth reading: How machine learning brings historical black and white photos back to life: How machine learning brings historical · How to transform your old black and white photos into vibrant memories with professional AI colorization: How to transform your old · Bring your vintage black and white wedding photos to life with AI colorization: Bring your vintage black and