What is video compression?
How codecs shrink your footage.

Raw video is enormous, so every practical codec is lossy. The decoded picture does not precisely match the source, and as Mozilla’s codec guide puts it, the more compression you achieve, the more detail and fidelity you lose. Compression itself is not the villain: your camera compresses, your NLE export compresses, and Apple sells ProRes 422 HQ, a compressed codec, as visually lossless, spending about 220 Mbps at 1080p to earn that label.

What hurts is the second, far harsher pass that happens after upload, tuned to a platform’s bandwidth bill instead of your picture. The rest of this page walks through what each stage discards.

Codecs almost never store a full picture for every frame. They store one complete image, an I-frame, self-contained like a JPEG, then describe the following frames as changes. P-frames borrow from the preceding I- or P-frame; B-frames borrow from both directions at once. Bitmovin’s worked example shows the payoff at 320 by 240: an I-frame averaged 14.8 KB on disk, a P-frame 6 KB, and a B-frame 765 bytes.

A run of frames starting with an I-frame is called a GOP, a group of pictures, and compressing whole GOPs instead of individual frames is where most of the bitrate savings come from. The cost is fragility: because each frame is built by applying changes to the previous one, errors compound over time, which is why a glitched stream smears and ghosts until the next full picture arrives.

Inside every frame, the encoder converts blocks of pixels into frequency coefficients and then quantizes them: fine detail gets rounded toward zero to save bits. Blocks with lots of variation resist this and stay expensive; flat, smooth areas compress readily. That budget logic decides exactly where a starved encode falls apart first.

Film grain is the worst case. Grain is random, so there is no spatial or temporal redundancy to exploit, and Netflix flatly calls it notoriously difficult to compress. AV1’s answer is to remove the grain before encoding, transmit its parameters, and resynthesize it on the viewer’s device: across the Netflix catalog, that trick alone cut average bitrates by 36 percent at 1080p and above.

Your eye resolves brightness far more finely than color, so codecs sample the two chroma channels at a lower rate than luma. In 4:4:4, every pixel keeps its own color sample. 4:2:2 halves the color resolution horizontally. 4:2:0, the format every consumer platform streams, halves it in both directions, keeping one color sample per four pixels.

For most shots the trick holds up. It fails at sharp color transitions: Netflix’s imaging team documented color bleeding and jagged edges from 4:2:0 exactly there, and found that with modern codecs, encoding full 4:4:4 is often the better deal anyway.

Bit depth sets how many shades each channel can express: 256 in 8-bit, 1,024 in 10-bit. HDR delivery is 10-bit territory by rule, not preference; YouTube requires 10- or 12-bit color for HDR uploads. When too few code values meet too-coarse quantization, smooth falloffs shear into visible bands, and encoders fight back with dithering, deliberately applied noise that randomizes the error and hides the steps.

Each codec generation buys the same picture for fewer bits, at the price of more decode work. The efficiency claims below are the published ones, not folklore.

Everything above happens twice: once in your export, under your control, and again after upload, when a platform re-encodes the file into its own bitrate ladder. The ceilings are public. Vimeo’s playback guidelines cap 4K H.264 at 22 Mbps. YouTube asks for 4K SDR uploads at 35 to 45 Mbps, then delivers 4K at about 13 Mbps by our measurement. Netflix, which controls its pipeline end to end, averages 8 Mbps on its top 4K rung.

The chart shows the gap. On uncompressed.io there is no ladder: the player streams the exact bytes you exported, so a 100 Mbps master plays at 100 Mbps in the client review room, and the full side-by-side lives on the Vimeo comparison page.

To go deeper: video bitrate, explained covers how many bits different delivery targets actually need, why your video looks worse after upload traces one file through a platform ladder artifact by artifact, and video hosting that does not compress is the buyer’s version of this page: how to verify a host leaves your file alone.