What the Copy Forgets

Most films are watched as a web copy: a file of a gigabyte or two that fits a whole feature into about two megabits a second. The same film also exists as a master, the version a disc or a 4K stream is made from, ten to thirty times larger. Everyone knows the small file is worse. This report asks what, exactly, is missing from it.

It is a narrow question asked carefully. Two films, Joker (2019) and Joker: Folie à Deux (2024), were each read from both copies, and for every one of their 2,622 analysed shots the same frame was taken from each. Then the master was turned into a web copy one step at a time, to see which step takes what. The answer is less obvious than "it is blurrier"; most of what you would expect to go survives, and the thing that changes most is not a loss at all.

The measurements concentrate on the dark end of the picture, because that is where these films live and where a small file is poorest. The companion report, The Colour of the Dark, found that both films have an unusual, cyan-leaning darkness; this one asks how much of that a web copy keeps.

Films. Joker (2019) · Joker: Folie à Deux (2024)

Web copies. 1080p x264, 8-bit, 2,100 and 1,341 kilobits a second

Masters. 2160p HEVC, 10-bit PQ HDR, 55.5 and 25.1 megabits a second

Shots. 1,302 and 1,320, the same frame in both copies

Region. the darkest part of the picture, L* 2–10, and everything above SDR white

Intervals. bootstrap over shots, 2,000 draws

1. method — A web copy has 16 levels to describe the darkest part of the picture; the masters have 61 and 65. In a typical Joker frame that is about 6,400 different dark colours against 28,000.
2. paired census — The web copy is not only a smaller master. Joker's shows the master's shadows exactly, on a 100-nit screen; Folie à Deux's was given its own shadow curve.
3. paired census + encoded sample — Grain is the main loss. Joker's web copy keeps about two thirds of the master's fine texture in the dark; the missing third goes to 1080p and to the encoder, in that order of size.
4. paired census — Everything above an ordinary screen's white is gone: half of Joker's shots and a quarter of Folie à Deux's have light the web copy cannot hold.
5. paired census + encoded sample — Compression barely touches colour. The colour changes between copies come from the second grade: Folie à Deux's web copy is yellower in the dark and its cyan turn is ten degrees shallower.
6. paired census — The look survives. Which colour channel empties first in the dark moves by under two points between master and web copy, in both films.

For 2,622 shots, the same frame was read from both copies of each film, and the web copies return exactly the numbers the earlier reports published.

Comparing two copies only means something if it is the same picture in both. The two editions are cut identically but start at slightly different moments: Joker's master runs 15 frames ahead of its web copy, Folie à Deux's not at all. At those offsets, 1,213 of Joker's 1,316 cuts and 1,258 of Folie à Deux's 1,352 land on exactly the same frame. Every shot was matched that way: 1,302 shots of one film, 1,320 of the other.

The second check is on the ruler rather than the pictures. The web copies were measured with the same code that produced The Colour of the Dark, and they give back its published numbers exactly: Joker red-first in 52% of its dark, Folie à Deux in 83%. So from here on, when a number moves between copies, the copy moved it. The measurement did not.

Figure 1 lets you see the whole comparison at once, before any of it is explained. Drag the line across the picture: the left side is the master, the right side the web copy, and the numbers underneath follow. Everything after it takes apart what you see there, starting with the simplest difference of all: how much each file is able to say about the dark.

Between near-black and dark grey, a web copy has 16 levels to work with and the masters 61 and 65. In a typical Joker frame that is about 6,400 different dark colours against 28,000.

A file stores every pixel as whole numbers, so it can only describe a range of darkness in so many steps. For the darkest part of the picture that still has light in it (the dark: lightness 2 to 10 on a scale of 100), an 8-bit web copy has 16 levels. The masters, with 10 bits and a brightness curve designed for dark detail, have 61 (Joker) and 65 (Folie à Deux) for the same range.

That sets the size of the smallest step. One level in the web copy's dark is a colour difference (ΔE) of about 0.7, where 1 is roughly the smallest difference you can see side by side. In the master one level is 0.23. The web copy's dark is a staircase with treads almost tall enough to notice; the master's is closer to a ramp.

The pictures use what they are given. A typical Joker frame from the master contains about 28,000 different dark colours; the same frame in the web copy has about 6,400. So far this is only capacity. A smooth dark wall needs few levels and a grainy one needs many, and which kind of dark these films have is the next question. Before it can be asked, one thing has to be cleared out of the way: the web copy is not simply the master made smaller.

Joker's web copy shows the master's shadows exactly as a 100-nit screen would. Folie à Deux's does not: it was given its own shadow curve, as if its master's white sat at 128 nits.

An HDR master is made for a screen that can shine at 1,000 nits or more; a web copy is made for an ordinary screen whose white is about 100. So a web copy of an HDR film comes from a second version, adjusted by hand or by a tool for the ordinary screen. Whatever that version changes was decided, and it has to be found before any loss can be counted.

It can be read from the pictures themselves. For each shot, the brightness of every pixel in the master was matched with the brightness of the same pixel in the web copy, level by level from black upward. For Joker the answer is plain. Up to about 10 nits, the web copy is the master shown on a 100-nit screen, one to one. Only above that does it start to fold the brighter light down, until 393 nits of master end up as the web copy's white.

Folie à Deux is different all the way down. Its web copy places the dark as if the master's white were at 128 nits, and the ratio slides as the picture gets brighter: a curve of its own. So some of what separates its two copies will be this second grade and some will be the file, and the pictures alone cannot say which is which. The only way to tell them apart is to make a web copy ourselves, from the master, with the losses and nothing else.

Made from the master with only the ordinary losses (1080p, 8 bits, halved colour and the web copies' own encoder settings), the rebuilt copies land within about 5% of the real ones on grain.

So the web copy was rebuilt from the master, one step at a time, and measured after every step:

• 1080p: the master shown at the brightness the real web copy implies, and reduced from 4K to 1080p.
• 8 bits: rounded to the web copy's 16 dark levels.
• Halved colour: colour stored at half the width and height of brightness (4:2:0).
• The encoder: squeezed by x264, the program that made both web copies, with each copy's own settings and its own bitrate.

These four steps are the ladder the rest of the report climbs down. Every step is applied to the same 2,622 frames, so each change is measured shot by shot, on identical pictures. The encoder step needs a second of film around each frame, so it was run on 200 shots per film.

It works. The rebuilt Joker copy has a grain of 0.58 against the real one's 0.55; Folie à Deux's rebuilt copy 0.51 against 0.52. So the ladder can be trusted to say which loss takes what, and whatever the rebuilt copy does not reproduce belongs to the second grade.

It very nearly said something false first. An early run found that the encoder turns the dark yellow. It does not: the tool feeding frames to the encoder was quietly converting their colours on the way in, because the frames were labelled on the way out but not on the way in. A lossless encode, one that must change nothing, changed things, and that is how it was caught. Every run now begins with that test and stops if it fails. The number that has to be zero is the most useful number in this report.

With a ladder that can be trusted, the report's question can be asked one loss at a time. The first thing it shows is where the texture goes.

Joker's web copy keeps about two thirds of its master's fine texture in the dark. The missing third is taken by 1080p and by the encoder; the colour grain goes when the colour is halved.

Grain is the fine, restless surface of a film image, and in the dark it is most of what there is to see. It was measured as how much each dark pixel stands out from the average of its neighbours. The Joker master scores 0.83, where 0 would be perfectly smooth; its web copy scores 0.55, about two thirds as much.

The ladder says where it goes. Showing the 4K master at 1080p takes 0.09, because four pixels become one and their differences average away. The encoder takes 0.14 more: to fit a feature film into two megabits a second, it treats fine, random texture as the most expendable thing in the frame, and in the dark there is little else. The 8-bit rounding in between takes nothing.

Colour grain, the flecks of green and magenta in a real dark, goes a different way. The master has 0.59 of it and the web copy 0.35. Most of that is lost when the colour is stored at half size, which the eye is meant not to notice. And 8 bits add some back that was never in the film: rounding a smooth dark to 16 levels scatters false colour flecks of its own (0.08 of them), which the encoder then partly smooths away. Part of what looks like texture in a web copy's dark is the file, not the film.

Folie à Deux loses much less, 0.59 to 0.52, and not because its web copy is better. Its master is a streaming file at 25 megabits, already compressed once, and it arrived with less grain to lose. The magnified details below show both films' dark lifted until it can be seen.

Grain is what the encoder takes. The next loss happens before the encoder ever runs: it is what the web copy's sixteen levels do to a dark that changes slowly.

What does the missing texture look like?

The same crop of the same frame, magnified 2×, with lightness 0–14 stretched to fill the whole range and colour multiplied by four so the dark can be seen. The master's dark is a field of grain; the web copy's is smoothed into soft blocks. These are analytical pictures, not what the film looks like.

Measured on: paired census.

At 8 bits, 3% of dark pixels have two colour channels exactly equal, which almost never happens in the master. Where the colour of the dark changes slowly, the web copy draws it in patches.

Sixteen levels show up in the pictures in two ways. The first is patches. Where a dark surface changes colour slowly, like a curtain's fold or a wall in half-light, the web copy cannot follow the change and draws it as flat islands, which the encoder then turns into blocks. In the red-curtain detail below, each pixel is painted by the colour channel closest to zero: smooth in the master, blotched in the web copy, with patches where green goes first that the master does not have.

The second is ties. At 8 bits, 3% of Joker's dark pixels have two channels exactly equal; in the master, with four times the levels, almost none do. That matters for any measurement that asks which channel is lowest, including the one The Colour of the Dark is built on: its counting gave ties to red, and that adds 1.3 points to Joker's red-first figure. It is a small error, and it is the web copy's.

The comb in Figure 1's histogram, the patches and the ties are not three findings. They are one fact about the file, sixteen levels, seen three ways, and that makes it a sturdier finding than any of the three alone.

Where does the web copy lose the fine changes?

Each dark pixel painted by its lowest colour channel — cyan where red is lowest, magenta where green is, amber where blue is — as deep as its saturation, for the same crop from both copies, magnified 2×.

Measured on: paired census.

Half of Joker's shots and a quarter of Folie à Deux's hold light brighter than an ordinary screen's white, up to 976 nits; the web copy folds all of it down. The masters' extra colours matter much less in the dark.

The largest thing the web copy does not have is not in the dark at all. The masters hold light far brighter than an ordinary screen can show: 46% of Joker's shots have highlights above 203 nits, the level HDR treats as an ordinary white, and its brightest reaches 976. For Folie à Deux it is 23% of shots, up to 785. In the web copy all of that is folded under 100 nits. Nothing there has been blurred or rounded. The web copy was never made to hold it.

The masters can also store colours outside the range an ordinary screen shows (called Rec.709). In the dark this matters less than it sounds. In a typical Folie à Deux shot, 36% of the darkest pixels lie outside the normal range (Joker: 7%), but only just: pulling them back inside moves 5.5% of them by a visible amount and 0.4% by a clear one. The master's cyan dark is a little beyond the web copy's reach, and only a little.

Every loss so far has been a loss of detail: texture, steps, range. None of them has changed the colour of the dark. That leaves the difference section 03 found and could not yet explain.

Compression leaves the colour of the dark almost exactly alone. What changes it is the second grade: Folie à Deux's web copy is yellower in the dark, its cyan turn ten degrees shallower and its vivid darks rarer.

If anything in a web copy were going to change a film's colour, the encoder would be the obvious suspect. It turns out not to be. Rebuilt from the master with the web copy's own settings, Joker's dark changes colour by less than a hundredth of the smallest visible difference, and Folie à Deux's by about a twentieth. The encoder takes texture and leaves colour, apart from a small cost in vivid darks: 1.6 points fewer of Folie à Deux's dark pixels stay fully saturated.

The colour differences between the copies come from the second grade. Folie à Deux's web copy moves its dark towards yellow (by 1.7 on the blue–yellow scale, b*). That takes about ten degrees off the cyan turn its dark makes against the rest of the picture, +51° in the master and +42° in the web copy, and 2.4 points off its vivid, fully saturated darks, 32% in the master against 26%. Joker's second grade does something different: it adds colour to the dark (+1.0 in chroma, how strongly coloured a pixel is) while barely changing which way it leans.

Neither was necessary to fit the file. They are what the film's makers, or the tool that made the SDR version, decided an ordinary screen should see. Which of the two it was cannot be told from the pictures, and for this question it does not matter: the web copy of Folie à Deux is a slightly different film in the dark, and its size has nothing to do with it.

Which colour channel empties first in the dark, the core of both films' look, moves by under two points between master and web copy.

After the texture, the fine steps, the light above white and a second grade, the look itself is still there. Joker's master is red-first in 55% of its dark and Folie à Deux's in 85.5%; the whole trip to the web copy moves each by 1.6 points, some of it the 8-bit ties and some the second grade. The finding of The Colour of the Dark, that these films colour their dark by emptying the red channel first, holds in the better copy, and is stronger there.

That is a smaller claim than "the web copy is fine" and a sturdier one. The web copy keeps what the film decided about colour; it loses what the film recorded about surface. How much surface there was to lose, though, depends on the film, and here the two differ.

Joker loses most of its difference to the file: grain. Folie à Deux loses most of its difference to the second grade: colour. Its master had already lost grain before this study began.

Put side by side, the two films lose different things. Joker's pair is a disc remux at 55.5 megabits a second and a web copy at 2,100 kilobits, 26 times less data. It differs mainly in texture, and the colour difference it does have is a web copy slightly more colourful in the dark. Folie à Deux's pair is a streaming master at 25.1 megabits and a web copy at 1,341 kilobits, 19 times less. It differs mainly in colour, because its web copy was graded differently, and less in texture, because its master had already been compressed once.

So the question "how much worse is the small copy?" has no single answer even for two films shot by the same cinematographer, graded by the same colourist and released five years apart. It depends on what the better copy is, and on who made the smaller one.

Watching a web copy of these films, you see their colour and their composition. You lose the surface of the dark, the fine steps between shadows and everything brighter than your screen's white. And for Folie à Deux you see a second grade.

The comparison can be said in one sentence per loss. The grain goes, mostly to the encoder. The steps between one dark and the next go, to 8 bits. The light above white goes, because the web copy was never made to hold it. And the colour of the dark does not go anywhere it was not sent: where it changes, someone made an SDR version that changed it.

These are statistics about still frames, and they are the form the loss takes here: not a verdict on whether anyone would notice, which only an audience could give, and not a description of motion, where compression does some of its most visible damage. What they do show is that "lower quality" is several different things, that they can be measured apart, and that the most interesting difference between these two copies of Folie à Deux is not a loss of information but a second opinion about its darkness.

Every technical term in the report, in plain words.

If a word above did not make sense, it is defined here, in the order you are likely to need it.

Every figure the prose quotes is declared with the path it came from in the measurements and re-checked when the report is built; a number that drifts stops the build. The encoder rung refuses to run unless a lossless encode through the same pipeline returns every frame exactly.

Three things would change the reading. A studio SDR master of either film would separate the second grade from anything that happened to the web copy on the way. Motion: measuring grain and blocking over time, not in single frames. And viewers: this report measures what differs between two files; it has not asked anyone which they prefer, or whether they can tell.

• Web copies: the 1080p x264 files measured in The Colour of the Dark; bitrates and encoder settings read from the streams.
• Masters: 2160p HEVC Main 10 PQ files (UHD remux for Joker, WEB-DL for Folie à Deux); mastering and light-level metadata read from the streams.
• Measurements: hdr_study/ — frames decoded by ffmpeg from the HEVC streams, never through an 8-bit still; the same toe, channel and band code as The Colour of the Dark.
• ITU-R BT.2408 (reference white 203 nits for PQ) · SMPTE ST 2084 (PQ) · ITU-R BT.709 and BT.2020.