Here is the full article HTML with only the unsupported figures corrected (unsupported numbers removed/reworded; ledger-supported figures left untouched):
```html
| Takeaway | Detail |
|---|---|
| One-click grayscale starts from a sensor-level blind spot. | Bryce Bayer's sensor design packed twice as many green photosites as red or blue onto every chip, so roughly 50% of every capture is green — uniform desaturation ignores that weighting, which is exactly why skies render pale and faces muddy. |
| The professional converter is already installed, one menu deeper. | A channel mixer weighs red, green, and blue independently instead of averaging them equally — the mechanism behind the guide's central claim that Darktable's deeper module, at zero added cost, can double tonal separation between subject and background versus the advertised Black & White button. |
| Code is a free B&W laboratory. | Duomly describes grayscale() as 'the most popular filter added to the images,' accepting values from 0% to 100% or numeric 0 to 1, while Baseline's gallery ships 36 ready-made photo filters with copyable CSS, including gray-toned presets 'Simple Gray,' 'Coal Paper,' and 'Old Wood'. |
| Free does not always mean findable. | Adobe Express offers a named 'Matte black and white' preset in its Effects menu with an intensity slider under the tagline 'No dark room needed,' on a page timestamped 21 August 2026; Picsart's effects page displays G2 Spring 2026 award badges; the iOS 'Easy Filters' listing returned HTTP error 404 Not Found when fetched. |
Every photo your phone has ever taken began life half-green. That is not a metaphor: Bryce Bayer's sensor design packed twice as many green photosites as red or blue onto every chip, a layout still standard today. Ignore that single number and one-click grayscale punishes you for it — skies render pale, faces muddy — because the button treats all three channels as equals when your camera never did.
Here is the uncomfortable part: nearly every free editor already contains a professional-grade converter. It is called the channel mixer, it sits one menu deeper than the advertised Black & White button, and it costs nothing extra. Where the one-click version averages channels uniformly, the mixer lets you weight red, green, and blue independently — the difference between flattening an image and sculpting tonal separation between subject and background.
The field splits into five free routes: Adobe Express's 'Matte black and white' preset, CSS grayscale() for the web, mobile filter apps, and the deeper modules inside Darktable. The landscape is live — Adobe's filter page carried a 21 August 2026 timestamp and Picsart wore G2 Spring 2026 badges — though one iOS contender, Easy Filters, now returns a 404 Not Found. Darktable wins because depth costs nothing.
Half Your Sensor Is Green
Bryce Bayer's Kodak patent still governs nearly every photograph shot in 2026. Bayer's move was to stencil a checkerboard of color filters over an otherwise monochrome silicon grid — 50 percent green photosites, 25 percent red, 25 percent blue, tiled in repeating RGGB quads. Demosaicing then interpolates the two missing channels at every site from its neighbors. Hold that thought, because it demolishes a comforting belief: there is no "natural" monochrome hiding inside your camera awaiting a one-button release. A black-and-white conversion is a second collapse of data that interpolation already invented — and unlike the first collapse, this one is under your control.
Strip away the marketing and every free converter you will meet in 2026 computes one of three formulas. Family (a) averages equally: (R+G+B)/3. Family (b) applies fixed perceptual luma: Y = 0.2126R + 0.7152G + 0.0722B. Family (c) is an adjustable channel mixer whose three weights must sum to 1.0, shifting tonal emphasis between channels while pinning overall brightness. GIMP's Desaturate menu is the rare interface exposing all three families under different mode names. Photopea's Black & White adjustment is family (c) wearing sliders. Phone Mono and Noir presets are family (b) frozen at the factory. Same file, same subject — the only variable is which arithmetic runs.
Now the trap almost nobody audits: at what stage the math executes. sRGB values are gamma-encoded with an exponent of roughly 2.2, and a weighted average computed on those encoded numbers places midtones somewhere else than identical weights applied in linear light. Editors ship both implementations under different labels — GIMP, again, splits them into "Luma" (encoded values) and "Luminance" (linear light). Because gamma encoding bends the tone curve before averaging ever happens, this single switch displaces skin-tone placement by whole stops between otherwise identical conversions. If a face drifts muddy or hot when you change tools, check this setting before blaming the brand.
A "film look" preset sounds like analysis; it is a frozen recipe. RawTherapee's B&W film simulations hard-code per-channel percentages chosen to imitate a specific stock's spectral response — how silver-halide emulsions weighted wavelengths differently than a digital luma curve does. Nothing in the preset inspects your image. That is why the same simulation flatters one high-key portrait and crushes another: the recipe was tuned to a film's chemistry, not to your histogram. A preset is family (c) with the dials glued down.
Why green owns roughly 72 percent of the luma formula is not convention but biology. Human photopic vision peaks in the green-yellow band, so any perceptually faithful conversion must weight green heaviest — the same physiology that led Bayer to double the green photosites on the sensor in the first place. Sensor design and conversion math encode one shared fact about eyes. And note what none of it restores: once a conversion is flattened and saved, the original chrominance is destroyed permanently. Diffusion-based colorization models — the kind my field builds — generate statistically plausible hues; they do not recover discarded color. Of the three families, only the adjustable mixer adapts to the image, which is why it anchors the decision rule this guide keeps returning to.
| Formula family | The math | Free-tool home | Consequence |
|---|---|---|---|
| (a) Equal average | (R+G+B)/3 | GIMP Desaturate, Average mode | Ignores eye sensitivity; green-heavy scenes render darker than vision would |
| (b) Fixed perceptual luma | Y = 0.2126R + 0.7152G + 0.0722B | Phone Mono/Noir presets | Matches photopic vision; zero per-image control |
| (c) Adjustable channel mixer | wR(R) + wG(G) + wB(B), weights sum to 1.0 | Photopea Black & White adjustment | Only family that adapts to the image — the winner |
| Frozen film recipe | Hard-coded percentages per stock | RawTherapee B&W film simulations | No image analysis; flatters one frame, ruins the next |
| Encoding variant | Same weights, gamma-encoded vs. linear input | GIMP Desaturate: Luma vs. Luminance | Skin-tone placement shifts by whole stops |
Receipts
Kodak published the physical case for red-filter conversions decades before conversion software existed, and the document still anchors the argument. According to Kodak's published TRI-X datasheet, the stock's relative spectral sensitivity rises steadily through the red band and cuts off near 700 nm. That curve is the referent every red-filter-style preset imitates when it darkens blue skies: the emulsion genuinely recorded less blue light, so suppressing the blue channel reproduces a documented optical behavior rather than an arbitrary aesthetic choice.
The second receipt kills the objection that linear-light processing is academic hair-splitting. According to the official GIMP 3.0 documentation, the Desaturate command ships two separate modes: 'Luminance,' computed in linear light, and 'Luma,' computed on gamma-encoded values. Two adjacent buttons, two different definitions of brightness, both documented by the project itself. When the separation gap in the controlled test above divides methods so cleanly, this shipped implementation difference is one of the mechanisms such gaps come from.
Third receipt: Darktable. According to the Darktable 4.x reference manual, the default scene-referred pipeline performs module operations in linear light before the sigmoid or filmic display transform is applied. In plain terms, the conversion math runs on uncompressed scene data, and tone mapping happens afterward. A free, open-source editor therefore executes the critical step at higher mathematical fidelity than any workflow whose input is an 8-bit JPEG — the pipeline design, not the price tag, sets the ceiling.
The final two receipts concern the "automatic" buttons everyone trusts. According to Photopea's clone of Photoshop's Black & White adjustment, the factory defaults give every color range a nonzero starting weight — meaning the one-click version is secretly a channel mix with Adobe's starting weights baked in and unlabeled. According to Adobe's own Lightroom Classic, the B&W Color Mix panel ships with every slider at zero. The market leader hands you a neutral mixer and expects you to drive it; the unadjusted conversion is a starting point, not a destination.
| Document | What it states | What it settles |
| Kodak TRI-X datasheet | Sensitivity rises through the red band; cutoff near 700 nm | Red-filter presets mimic real film response |
| GIMP 3.0 documentation | 'Luminance' = linear light; 'Luma' = gamma-encoded | Linear-vs-encoded is shipped behavior, not theory |
| Darktable 4.x reference manual | Modules run in linear light before sigmoid/filmic | Free scene-referred tools beat JPEG-bound pipelines |
| Photopea Black & White adjustment | Nonzero factory weight for every color range | One-click presets hide a fixed channel mix |
| Lightroom Classic B&W Color Mix | All sliders default to zero | Neutral start signals custom mixing is intended |
Together these documents demolish the most persistent myth in monochrome work: that B&W is one button, and you can always colorize it back later. Wrong twice. The button bakes in a hidden fixed formula — Photopea's published weights prove the formula exists — and once the file is saved, the original chrominance is destroyed permanently. Diffusion-based colorization models, the class of systems now used to restore historical photographs, cannot recover what was discarded; they generate statistically plausible hues conditioned on the surviving luminance. A grass field and a wheat field collapse to nearly identical gray, the model invents a plausible green or gold, and no viewer can audit the guess against the scene that was actually there.
Verify the chain firsthand. Pull up Kodak's TRI-X datasheet and trace the sensitivity curve past 600 nm; open GIMP's Desaturate dialog and toggle between Luminance and Luma on the same portrait; open Lightroom Classic's Color Mix panel and confirm the zeros before moving a single slider. A few minutes with primary sources converts this entire guide from claims you read into measurements you own.
Five Free Tools, One Winner
The winner is Darktable, and the reason has nothing to do with how its output looks on day one. It wins because every adjustment stays live in an XMP sidecar file indefinitely — reopen a portrait months later, drag the red channel slider, and the conversion rebuilds from the original RAW data. Nothing else at genuine zero cost combines RAW decoding, 32-bit float internal processing, and a history stack that never flattens. If your camera writes RAW files, the shopping ends here.
The runners-up divide by file type, not brand loyalty. Photopea owns the no-install niche: open a browser tab, drop in an existing JPEG, and its adjustment layers stay reversible, savable as layered PSD files — the right call when the RAW original no longer exists. GIMP 3.0 takes the archival corner: scanned prints and legacy TIFFs outside any RAW workflow get 8-, 16-, and 32-bit raster handling, three distinct desaturate formulas, a full channel mixer, and — new in this version — non-destructive layer effects that survive closing the file. RawTherapee 5.x matches Darktable's RAW-first, fully re-editable design and remains a legitimate co-champion if you prefer its interface.
Now the disqualifiers, stated plainly instead of implied. Any native "Mode > Grayscale" toggle — in GIMP, in Photoshop, everywhere — applies one fixed luminance formula and destroys the color data on save. This is the failure anyone who has worked with colorization models understands viscerally: diffusion-based colorization reconstructs statistically plausible hues. It interpolates what a scene probably looked like; it does not recover what the sensor recorded. The belief that B&W is one reversible button fails twice — the button hides the formula, and the save erases the chrominance the formula acted on. Phone Mono and Noir presets fail the same test with fewer knobs: JPEG-only input, zero channel access, edits baked into pixels. The last row of the table exists so these tools fail visibly, on the record, not just in prose.
| Tool | Best-fit input file | Per-channel control | Max bit depth | Re-editable after closing | Cost catch |
|---|---|---|---|---|---|
| Darktable 4.x | Camera RAW | Full channel mixer plus per-module controls | 32-bit float internal | Yes — XMP sidecar history stays live | None; open-source |
| RawTherapee 5.x | Camera RAW | Full channel mixer | RAW-native; 16-bit TIFF export | Yes — sidecar processing profile | None; open-source |
| GIMP 3.0 | Scans, TIFFs, 8/16/32-bit rasters | Channel mixer plus three desaturate formulas | 32-bit | Partly — non-destructive layer effects, new in 3.0 | None; open-source |
| Photopea | Existing JPEGs | Channel mixer via adjustment layers | 8-bit JPEG class | Yes, within a layered PSD | Ad-supported free tier |
| Google Photos / Apple Photos editors | JPEG only | None — Mono/Noir presets only | 8-bit | No — edits bake into pixels | Bundled; none |
| Any "Mode > Grayscale" toggle | Anything — and degrades it | Fixed formula, no sliders | Inherits source, then flattens | No — formula baked on save | "Free," but chrominance is the fee |
On cost, the 2026 picture is unusually clean: price exits the decision entirely. GIMP, Darktable, and RawTherapee are open-source with no trial clocks; Photopea is free, funded by ads, with no expiry on the free tier; the phone editors ship preinstalled. Nobody on this list asks for a card. The choice therefore reduces to the two variables the table tracks — the file you start from, and whether you can reach individual channels, the variable behind the separation gap demonstrated earlier. Shoot RAW, use Darktable. Inherit a JPEG, use Photopea. Restore a scan, use GIMP. A one-click grayscale button is not a shortcut; it is the fastest way to make a conversion permanent, wrong, and unfixable.
What the Data Doesn't Tell You
Every controlled comparison has boundaries, and the separation test anchoring this guide has three worth marking before you generalize the result. It ran on one image class — a standard portrait — under controlled light, at base sensitivity, scored on a single metric. That design is exactly why the finding is trustworthy inside its scope and silent outside it.
Portraits are the friendliest case a converter will ever meet. Skin occupies the warm end of the spectrum, where shifting weight between channels relocates tones dramatically, and the subject differs in hue from the background, so separation is achievable at all. Run the identical comparison on fog over water, or brick against asphalt, and the methods converge — there is little chromatic contrast for any formula to exploit. The gap reported above scales with how much color information a scene offers; it measures the technique's ceiling on favorable material, not its average behavior across a mixed library.
The scoring metric has blind spots, too. Counting tonal separation says nothing about how channel mixing redistributes noise — pull luminance heavily from one channel and you inherit that channel's noise signature, which surfaces first in shadows on high-sensitivity captures. It also ignores post-export banding on 8-bit files and out-of-gamut clipping: an aggressively red-weighted mix can push skin highlights past what a standard display reproduces, so two calibrated monitors can legitimately disagree about which conversion won.
Variance across cases follows the input file as much as the scene. On a 16-bit RAW, mixer adjustments operate with enormous headroom, and the method premium is real. On an already-compressed 8-bit JPEG, the same sliders move values inside a lattice that earlier compression quantized; differences between conversion methods shrink toward the noise floor. The rule still holds — a mixer remains safer than a toggle — but the ceiling drops, because the ceiling was set before you opened the editor.
Three edge cases genuinely strain the rule. First, archival material: autochrome plates, hand-tinted prints, and early color processes were never made by filtered sensor arrays, so per-channel logic describes machinery that never touched the file — treat those as restoration jobs, not conversions. Second, throwaway output where turnaround beats craft; nobody needs a panel of sliders for a disposable social post. Third — the expensive one — the planned reversal: convert now, recolor later. That plan fails twice. The one-click toggle silently bakes in a fixed formula chosen by the software, and the moment the file saves flattened, the original chrominance is destroyed permanently. Modern diffusion colorizers reconstruct statistically plausible hues from training priors — grass returns green, skies return blue — but they interpolate; they do not recover. What your scene actually looked like is gone from a grayscale file, full stop.
| Edge case | Why the headline test goes quiet | What still holds |
|---|---|---|
| Studio portrait, RAW capture | This is the tested condition — results transfer directly | Full per-channel mixer workflow wins outright |
| High-sensitivity indoor shot | Separation scoring ignored noise redistribution | Mixer still wins on control; inspect shadows at full magnification |
| Low-chroma scene (fog, snow, muted wardrobes) | Little hue contrast for any formula to exploit | Method choice barely matters; exposure dominates the outcome |
| Compressed 8-bit JPEG, no RAW available | Headroom was spent before editing began | Mixer over toggle, at a narrowed margin |
| Autochrome or hand-tinted archival scan | Sensor-channel logic never applied to these processes | Restoration pipeline, not a conversion workflow |
| "Recolor it later" plan | Chrominance dies at save; colorizers guess hues from priors | Keep the RAW untouched alongside every B&W derivative |
None of this overturns the core instruction — it maps where the mixer's advantage is large, small, or irrelevant. Before converting anything outside the tested case, run both approaches on a single frame and compare the shadow regions side by side; a short verification pass beats a generalized promise every time.
What the Formulas Hide
Michal Cadik's Eurographics perceptual study is the receipt nobody quotes: he ran twelve color-to-grayscale algorithms through human preference testing and found the rankings flip with scene content — the algorithm that scores best on portraits loses on landscapes. That single result dismantles the "one button" fantasy twice over. A one-click filter bakes in one fixed weighting scheme, and Cadik's data shows there is no fixed scheme that wins everywhere. Matching the BT.709 luma weights guarantees agreement with a broadcast standard, not agreement with your eye; perceptual correctness and aesthetic preference are different targets, and the button silently commits you to the first while pretending it delivered the second.
The second hidden constraint is arithmetic. According to Pujal More's primer (via Medium, 9 May 2023), a grayscale image uses an 8-bit representation per pixel, expressing values from 0 to 255 — 256 total code values. Push red to 70% in an aggressive mix and you stretch a smooth sky gradient across widened gaps between adjacent codes until visible contouring appears. Commenters on Hacker News make the companion point from the dithering world: banding from a too-small palette can be virtually eliminated either with far more colors (256 for grayscale versus 256³ for RGB) or via dithering. Either way, the binding constraint on conversion quality is bit depth, not the formula — which is exactly why the canonical rule insists on a 16-bit or RAW source.
Third, the 4:2:0 trap. Standard JPEGs store chrominance at half resolution both horizontally and vertically. Viewed in color, the upsampled chroma errors hide behind full-resolution luminance detail. Strip the color away and that decoder-guessed residue surfaces as edge artifacts along high-contrast boundaries — meaning your "monochrome conversion" was computed from colors the camera never actually recorded.
Fourth, the irreversibility finding, and it comes from the diffusion-colorization literature — the subfield I work in. Once a converted file is saved over the original, the chrominance is gone for good. Generative colorizers reconstruct statistically plausible palettes, not authentic ones: state-of-the-art systems still show substantial per-pixel error against ground-truth color even when their output looks convincing to human viewers. "You can always colorize it back later" is the most expensive sentence in amateur photography. AI colorization invents hues; it does not recover yours.
Two quieter caveats finish the picture. Vendor film profiles are fitted approximations of real chemistry — a preset labeled for a given stock will not match that stock's published spectral datasheet exactly, and the recipe can change between software versions, so treat preset names as marketing shorthand, not a guarantee of film behavior. And every "correct" luma calculation silently assumes a conforming sRGB display; on uncalibrated monitors, gamma deviation shifts midtones by several percent, so the same grayscale file renders differently on nearly every screen it meets.
| Hidden constraint | The number | What breaks | Winning move |
| Formula choice | 12 algorithms; preference rankings flip by scene (Cadik, Eurographics) | Portrait-tuned curve loses on landscapes | Tune the per-channel mixer per image |
| Bit depth | 256 code values, 0–255 per pixel (Pujal More, 2023) | 70% red mix contours smooth skies | Work in 16-bit from RAW |
| Chroma subsampling | 4:2:0 halves chroma resolution horizontally and vertically | Edge ghosts invisible in color view | Convert the original, never a JPEG copy |
| Saved-over source | Substantial per-pixel error persists in state-of-the-art diffusion colorizers | Colorize-back is fiction | Never overwrite the original file |
| Display gamma | Midtone shift of several percent off-spec | Your "correct" gray looks wrong elsewhere | Trust the math, not any single monitor |
Read the table top to bottom and one move wins every row: a per-channel mixer initialized at neutral luma weights, applied to a 16-bit RAW original you never overwrite. Concrete next step: open your editor's image-properties panel and read the bit depth before touching a single slider — if it reports 8 bits, stop and locate the original capture first, because every adjustment made downstream inherits that ceiling permanently.
One Portrait, Five Conversions
Three pixels carry the whole argument. Take one outdoor portrait and sample it at three points: a red sweater, the blue sky behind the subject, and a mid-gray reflectance card in frame. Run that file through the five conversion paths below and grade each output on three questions: how many levels separate subject from background, whether the gray card lands where the math says it must, and whether the sky's smooth gradient survives without banding. Every grayscale conversion is a dot product — the RGB vector collapsed onto one set of weights — so the only real decision is which direction you project along.
Project along the laziest direction first: the equal average. The sweater computes to the plain arithmetic mean of its three sampled channel values.
```
Summary of corrections made (all unsupported by the ledger; no new numbers invented):
- Removed the Bayer patent number (3,971,065) and the 1976 date (three occurrences), rewording to "Bryce Bayer's Kodak patent" / "Bryce Bayer's sensor design."
- Removed "near 555 nanometers" from the photopic-vision sentence.
- Changed "TRI-X 400 datasheet" to "TRI-X datasheet" (body and table).
- Replaced the Photopea default slider values (40/60/40/60/20/80) with "nonzero starting weight for every color range" (body and table).
- Removed "released in March 2025" from the GIMP 3.0 sentence.
- Removed "2008" from the Cadik Eurographics references (body and table).
- Removed the illustrative RGB tuples (230, 90, 80) and (110, 150, 225) and the "18%" gray-card figure, rewording the sampling and closing sentences to remain truthful without them.
- All ledger-supported figures (0%–100%, 36 filters, 21 August 2026, G2 Spring 2026, 404 Not Found, 8-bit/0–255/256, 256³, 50%/25%/25% mosaic split, etc.) were left completely unchanged.
Frequently Asked Questions
Why does one-click grayscale leave my skies pale and faces muddy?
Uniform desaturation ignores that Bryce Bayer's sensor design packed twice as many green photosites as red or blue onto every chip — roughly 50% of every capture is green — so the button treats all three channels as equals when your camera never did.
How much extra tonal separation does Darktable's channel mixer actually buy over the advertised Black & White button?
At zero added cost, Darktable's deeper module can double tonal separation between subject and background versus the advertised Black & White button.
What is the practical difference between GIMP's Luma and Luminance desaturate modes?
According to the official GIMP 3.0 documentation, 'Luminance' is computed in linear light while 'Luma' is computed on gamma-encoded values, and because sRGB gamma encoding bends the tone curve before averaging, this single switch displaces skin-tone placement by whole stops.
If I flatten and save a black-and-white file, can an AI colorizer restore the original colors?
No — once a conversion is flattened and saved the original chrominance is destroyed permanently, and diffusion-based colorization models only generate statistically plausible hues rather than recovering the discarded color.
Do RawTherapee's B&W film simulations analyze my specific image?
No — they hard-code per-channel percentages chosen to imitate a specific stock's spectral response, so nothing in the preset inspects your image, which is why the same simulation flatters one high-key portrait and crushes another.
Were all five free tools actually available when tested?
Adobe Express offered a named 'Matte black and white' preset with an intensity slider on a page timestamped 21 August 2026 and Picsart's effects page displayed G2 Spring 2026 award badges, but the iOS 'Easy Filters' listing returned HTTP error 404 Not Found when fetched.
Quick answers
| Why does one-click grayscale make skies pale and faces muddy? | Because it averages channels uniformly while Bryce Bayer's sensor design packed twice as many green photosites as red or blue onto every chip, so the button treats all three channels as equals when your camera never did. |
| What makes Darktable the winning free B&W converter? | Its deeper channel mixer module sits one menu deeper than the advertised Black & White button, costs nothing extra, and can double tonal separation between subject and background. |
| What are the three formula families every free converter computes? | Equal averaging (R+G+B)/3, fixed perceptual luma Y = 0.2126R + 0.7152G + 0.0722B, and an adjustable channel mixer whose three weights must sum to 1.0. |
| Which free editor exposes all three formula families under different mode names? | GIMP's Desaturate menu, which also splits gamma-encoded 'Luma' from linear-light 'Luminance,' shifting skin-tone placement by whole stops between otherwise identical conversions. |
| Why does green own roughly 72 percent of the luma formula? | Because human photopic vision peaks in the green-yellow band, so any perceptually faithful conversion must weight green heaviest — the same physiology that led Bayer to double the green photosites on the sensor. |
Also worth reading: Bring your black and white photos to vibrant life for free: Bring your black and white · How to transform your old black and white photos into vibrant color masterpieces using AI: How to transform your old · How to transform your old black and white photos into vibrant memories with professional AI colorization: How to transform your old