Precision of values. Density of samples.

An n-bit integer component has 2ⁿ possible codes: 256 at 8 bits, 1024 at 10 and 4096 at 12. This is the whole word, not automatically the number of steps between nominal black and white. Video mapping uses part of the scale differently, and luma and chroma spans differ.

Rounding replaces a continuous or more precise value with a permitted level. For a fixed transfer and range, a finer grid reduces error. Yet “8 versus 10” is incomplete without the function, range and earlier losses. Storing eight-bit-derived values in a ten-bit container does not retroactively create precision.

When neighbouring regions round to the same code, a continuous change becomes a plateau. Strong later grading can make boundaries more visible. Banding is not caused only by source bit depth: compression, processing, export, the graphics path and display also matter. A screenshot alone cannot identify the culprit.

Small noise or dither can decorrelate quantisation error and make its structure less conspicuous; it does not reconstruct missing original values. The lab intentionally omits dither to expose rounding first. Its enlarged ramp explicitly stretches 0.25…0.35; it is not HDR viewing.

Measurement dynamic range concerns useful upper signal relative to a lower limit determined by noise and the chosen criterion. Recording bit depth specifies representation precision. A nonlinear function can allocate a large light range to a limited number of codes without improving physical signal-to-noise ratio.

There is no universal conversion from “10 bits” to a fixed camera stop count. More codes may preserve distinctions better within one range; another curve distributes the same number differently. Ask separately what was measured, which range was encoded and how precisely each value was stored.

Processing may use floating-point values, allowing negative intermediates and values above one without immediate clipping. That is neither infinite precision nor a guarantee that every plug-in, LUT or export retains the same range.

Converting to float does not restore discarded chroma or undo earlier rounding. It helps avoid extra limitations during subsequent operations. Check pipeline boundaries separately: decoding, interchange, effects, external rendering and final encoding. One project setting does not certify the whole chain.

In the Y′CbCr representation considered here, 4:4:4 retains Y′, Cb and Cr at every position. It preserves a complete chroma grid without mandatory spatial reduction. It does not by itself specify depth, codec, compression strength or transparency.

For W×H imagery, three full planes give 3WH samples excluding alpha. That counts spatial values, not compressed file size. ProRes profiles have different capabilities; 4444 includes alpha support but does not mean every such file actually contains meaningful transparency.

Typical 4:2:2 halves Cb and Cr density horizontally while retaining vertical density. Each pair of luma positions in a row has one Cb and one Cr sample. It does not delete half the RGB pixels or divide the image into red, green and blue portions.

For even W and H, the total is 2WH values: WH luma and WH/2 per chroma component. At equal depth, that is two thirds of the 4:4:4 sample count before compression. Fine vertical colour boundaries are sensitive to horizontal filtering, while grey edges can retain detail through Y′.

In a usual progressive 4:2:0 model, one Cb/Cr pair serves a 2×2 luma region. Y′ stays full resolution; each chroma plane has WH/4 samples, for a 1.5WH total. That is half the total component-sample count of 4:4:4, not “half the image quality”.

Colour differences are filtered before reduction. The lab uses box averaging followed by repeated-block reconstruction. Real codecs and decoders can use different filters and positions. The preview explains a class of loss; it does not predict the quality of a particular H.264 or HEVC preset.

4:2:0 alone does not fully specify chroma positions relative to Y′ or reconstruction. Progressive and interlaced arrangements deserve separate attention. Filter phase and sample positions influence fine colour edges; a mismatch can look like a halo or colour shift.

Shifting the lab pattern shows how identical structure aligns differently with averaging blocks. It is not an MPEG chroma-siting selector: the model positions are fixed and documented. For real interchange, inspect chroma_location metadata and the resampler behaviour.

Each export can add quantisation, spatial filtering or codec loss. Yet “every repetition must cause more damage” is too broad. Repeating the same already-rounded operation can be idempotent; real loss depends on new values, filters, phases and compression.

Risk is greater when size, alignment, grading or formats change between exports. Converting 4:2:0 back to 4:4:4 reconstructs a dense grid, not the original coloured detail. Such an intermediate can be valuable for later processing, but it is not a time machine.

Smooth gradients and heavy grading need value precision. Fine coloured graphics, chroma-key boundaries and intermediate compositing also need spatial sampling. Motion, texture, noise and codec limitations add other requirements. There is no universal one-word ranking of 10-bit 4:2:0 against 8-bit 4:4:4.

Start from the most vulnerable element and run a short test, then consider compatibility, storage and recipient requirements. Footage that already lost chroma during capture can still benefit from a careful intermediate preserving newly added titles and subsequent calculations.

Use a short test with sky or a smooth fill, fine coloured lettering and a moving edge. Compare source, processed result and re-imported export under matching colour management. Viewer scale matters: downscaling can hide defects, while enlargement can emphasise interpolation structure.

Inspect the actual output pixel format and profile separately. Render settings describe intent; the encoded stream is what the recipient will read. For keyed material, inspect both RGB and matte because chroma-edge errors may become more visible after selection.

A working master should retain properties needed for later processing; a delivery file should meet the recipient’s confirmed specification. These may be different files. Do not use one universal “maximum” preset: an unsupported profile is not made useful by a higher bit count.

Check component depth, chroma, range, primaries, transfer and alpha separately. Retain sources and processing notes. If a platform transcodes the file, inspect its actual result as well as the local copy. Confirm platform requirements when delivering.