RGB Channel Splitter

A colour image stores each pixel as a few numbers. In an RGB image those are three whole numbers from 0 to 255 — how much Red, Green and Blue the pixel contains — and many formats add a fourth, Alpha, for transparency. A channel is one of those numbers taken for every pixel and drawn as its own picture. Splitting an image into channels answers a simple question: what does each of those numbers look like on its own?

This page does four things with them. It splits an image into labelled channel tiles with simple statistics; it inverts a channel (255 minus each value) for inspection; it swaps channels, for example putting the Blue values into the Red slot; and it combines channels, taken from the loaded image or from separate files, back into one RGBA image. Split, inspect, swap and recombine all happen in one place, and the image never leaves your browser.

Everything here works on whole 0–255 values. It never looks inside a value's individual bits (the Bit Plane Viewer does that), it never hides or searches for a secret message (the Color Channel steganography tools do that), and it does not draw histograms. If a channel looks suspicious, the Next steps tab points to the tools that dig deeper.

A channel is a single number per pixel, so by default each tile is drawn as a gray picture: 0 is black and 255 is white. That is why the Red tile of a red-and-blue photograph looks like a black-and-white version of it: bright where the original had a lot of red, dark where it had none. Switch the display to In channel color to tint the tile instead (Red as value, 0, 0; Green as 0, value, 0; Blue as 0, 0, value). The tint is only a different way of drawing the same numbers.

Alpha has no colour of its own, so its tile is always gray: white means fully opaque and black means fully transparent. Gray and the derived CMYK, HSV and HSL channels are also always gray. Every tile carries a text label (Red (R), Green (G), Blue (B), Alpha (A) or Gray), so you never have to rely on tint alone to tell the tiles apart.

Under each tile you see the smallest value, the largest value and the mean, with a note saying whether they describe the Original or the Inverted values. The sample image below shows how this works: its Red channel is 255, 0, 0, 255, 10, 255 for the six pixels, so min 0, max 255, mean 129.1667; inverting turns each value v into 255 − v and the mean into exactly 125.8333.

A useful habit: look at a channel at full size, then at its inverted version. Large flat areas, a pattern that is faint in one channel but strong in another, or a shape that appears in only one channel are all worth a closer look. They are leads, not proof of anything hidden: photographs often have uneven channels simply because of lighting and colour.

The tiles shown depend on how the file stores its colour, and the Load tab says which and why. A plain grayscale PNG stores one value per pixel, so you get a single Gray tile rather than three identical ones. A gray-plus-alpha PNG gives Gray and Alpha. RGB and RGBA PNGs give Red, Green, Blue (and Alpha). A palette image (PNG with a colour table, an 8-bit BMP, a GIF) stores a colour-table index per pixel; the tool looks each index up in the table, so the Red, Green and Blue tiles hold the colours the table actually contains, never the index number itself.

Transparency depends on the type. A PNG with a tRNS chunk behaves in one of two ways. In a palette PNG the chunk lists an alpha value for each colour-table entry, so the tool shows a real Alpha tile built from those values, even when every pixel happens to be fully opaque. In a grayscale or RGB PNG the chunk names a single transparent colour, a binary key and not an alpha channel, so no Alpha tile is offered; the Load tab reports the key and how many pixels match it. The match is made on the sample stored in the file at its own bit depth, before any scaling.

A 32-bit BMP has a fourth byte per pixel that the format reserves and does not define as alpha. This tool never uses it as transparency, including in Combine; a Show reserved byte switch lets you look at it, labelled Reserved byte (not alpha). JPEG has no transparency at all. For WebP and AVIF an Alpha tile appears when the decoded image has transparent or partly transparent pixels, or when the file's own header declares an alpha channel. For a GIF only the first frame is used. Pixels that frame does not paint, whether because of a transparent colour index or because they lie outside the frame's rectangle, count as transparent and give an Alpha tile. That is how Chromium, Firefox and WebKit all drew a test GIF whose frame was smaller than its screen; the GIF's logical-screen background colour is not used.

Bit depths below 8 are expanded to 0–255 by exact whole-number factors: 1-bit gray multiplies by 255, 2-bit by 85 and 4-bit by 17, so no rounding ever happens. A 16-bit PNG shows the high byte as the 8-bit value and keeps the full 16-bit samples; a switch on the Split tab makes the statistics, the inverted values and the grayscale downloads (single channel, ZIP and pixel table) use those full samples, and Combine has its own Output precision setting.

Combine has four slots: Red, Green, Blue and Alpha. Each slot can be empty, can take a channel of the loaded image, or can take a separate uploaded image. This is the case where a challenge hands you three grayscale files and says each is one channel of the real picture: upload them into the three slots and combine. You can also pick Inverted values for any slot, which feeds 255 minus each value.

What a slot reads from an uploaded image is fixed, not guessed. For the Red, Green or Blue slot, a grayscale or gray-plus-alpha image supplies its gray value, and an RGB, RGBA or palette image supplies its own Red channel (no averaging or brightness blend). For the Alpha slot, a grayscale image supplies its gray value, and an image with real alpha supplies that alpha; an image with no alpha channel is refused for that slot with a message saying so. One Gray source can fill the Red, Green and Blue slots together to give a neutral gray result.

The first filled slot, in the order Red, Green, Blue, Alpha, sets the output size, and every other filled slot must match it exactly. A mismatch is refused with both sizes and the slot named; nothing is resized or cropped. The size is shown before you press Combine, and Combine stays disabled until at least one slot is filled. Empty Red, Green and Blue slots become 0 and an empty Alpha slot becomes 255 (fully opaque).

The result is always an RGBA PNG, even when Alpha is empty. Swap is a shortcut on top of the same machinery: it fills the slots with the loaded image's own channels in a different order and lets you combine at once. Swapping Red and Blue turns the pixel (10, 20, 30) into (30, 20, 10).

A channel can be downloaded as a grayscale PNG, in which the three colour values all equal the channel value. This is the version meant for re-use: drop it back into a Combine slot and the slot reads exactly the value you exported. It always holds the values you have selected (original or inverted) and never depends on the display mode that was on screen.

The tinted preview is a separate button. It bakes the hue into the picture (Green as 0, value, 0), which is nice to look at or share, but the file no longer contains the channel as a value a Combine slot can read: an RGB image supplies its Red channel to a slot, and the Red byte of a tinted Green picture is 0. For that reason the tinted preview is labelled presentation only, and the ZIP (one PNG per channel, named like photo-red.png, with -1, -2 added if two names would collide) always contains the grayscale version.

The contact sheet puts the Original and every channel tile in one labelled strip. It is a presentation image: its tiles are scaled to fit a capped canvas (at most 8000 by 3000 pixels), and a transparent original is drawn over a checkerboard instead of being silently flattened. The pixel table gives one row per pixel (x, y and each channel at the selected values) as CSV or TXT, up to 1,000,000 pixels. JPG and WebP exports of a channel preview or of the combined image are lossy and are never meant to be re-uploaded; JPG has no transparency, so it is flattened onto white.

For colour images the Split tab can show CMYK, HSV or HSL instead of RGB. These are derived from the 8-bit Red, Green and Blue values with fixed formulas and scaled to 0–255 so they can be drawn, inverted, measured and exported like any channel. Key is 255 − max(R, G, B); Cyan, Magenta and Yellow are (max − channel) × 255 ÷ max, rounded (for pure black, where max is 0, C, M and Y are 0 and K is 255); Hue is its angle × 255 ÷ 360; Saturation and Value (or Lightness) are scaled to 0–255. Rounding is always half up, in whole-number arithmetic. These are analysis views: Combine still takes Red, Green, Blue and Alpha only.

Compare takes a second image of the same size and shows, per channel both images have, the absolute difference, how many pixels differ, the largest difference and the mean difference. A gray image counts as Red = Green = Blue. Amplifying makes small differences visible on screen; the numbers and the downloadable difference file always carry the real values. It is a way to see where two near-identical pictures disagree, for example an original and a re-saved copy.

For a 16-bit PNG, the full-precision switch on the Split tab changes the statistics, the inverted values (65535 − v), the grayscale downloads (single channel and ZIP, written as 16-bit PNGs) and the pixel table to use the complete samples. Combine is separate: its Output precision setting is 8-bit by default and offers 16-bit when a 16-bit source is in use, which writes a 16-bit RGBA PNG (full samples from 16-bit sources, 8-bit sources scaled by 257, inversion as 65535 − v). Everything that is drawn or converted for looking at stays 8-bit: the tiles on screen, the tinted preview, the contact sheet and the JPG/WebP previews.

PNG and BMP files (and the first frame of a GIF) are decoded by code written for this page. For PNG that means validating every chunk checksum, undoing the scanline filters, resolving palettes and reading the samples directly, so the values equal what is stored in the file. This is called exact sample extraction. A file that is malformed is refused with a specific message naming what failed (for example a bad checksum, a truncated image-data chunk, a palette index beyond the table or an unknown critical chunk), not partially drawn.

JPEG, WebP and AVIF are handed to your browser's image decoder first. The tool asks the browser to switch off orientation handling, colour conversion and alpha premultiplication where it allows; in our test Chromium, Firefox and WebKit all still applied a JPEG's EXIF orientation. That is browser-decoded extraction: values can differ from the file because of colour management or resampling, JPEG is lossy by nature, and a banner says so, naming any option the browser would not switch off. The size of the difference depends on the browser and the file. In a test on seven real photographs (six of 356 × 356 pixels and one of 1000 × 1000) saved as JPEG at quality 95 and 75 and as WebP at quality 90, the average difference from the exact PNG was between 0.6 and 1.9 levels (out of 255) in Chromium and Firefox, and a lossless WebP matched exactly in all three browsers tested. In WebKit (the engine behind Safari) the same WebP files matched Chromium, but the JPEGs, which carried an embedded colour profile, averaged about 4 levels (1.3 to 9.3 depending on the photograph) and differed by up to 86 for a single value, because that browser applied the profile. When a JPEG or WebP embeds a colour profile the tool says so in the banner. A JPEG's EXIF orientation tag is another case: Chromium, Firefox and WebKit all applied it in our test even though the tool asks them not to, so such a JPEG is shown rotated and the banner says so. The conditions of the measurements are listed in the table below. When the result matters, work from a lossless original such as PNG.

Typical jobs: checking whether one channel looks different from the others, reassembling a picture from three or four separate channel files, trying the six Red/Green/Blue orderings when a picture's colours look wrong, and isolating one channel for a mask or a colour-cast diagnosis. Look at the tiles in grayscale first, then inverted, then expand any that look different and read pixel values with the inspector.

What to watch for, and what is not evidence: a channel that is nearly flat while the others are busy, a faint outline visible in only one channel, or a swapped ordering that suddenly makes a picture natural. Smooth gradients, saturated colours and shadows make channels differ in ordinary photographs, and an alpha channel that is uniform is normal. This page does not decide that anything is hidden; it shows channels. Statistical or bit-level evidence needs the Steganography Detector or the Bit Plane Viewer, and recovering a message needs an extractor.