Precision and detail
What 8/10/12-bit and 4:4:4/4:2:2/4:2:0 actually provide: quantisation, banding, grading headroom, spatial chroma grids, siting, keying and generation loss.
Bit depth controls the precision of each sample value; chroma subsampling controls how often colour-difference samples exist spatially. 10-bit 4:2:0 and 8-bit 4:4:4 can each be preferable for different tasks.
Quick reference
4:2:0 also requires a location/filtering convention.
PRECISION × SPATIAL CHROMA
In a full-range representation, n bits provide 2^n code values: 8-bit gives 256, 10-bit 1024 and 12-bit 4096. Narrow/video range uses a subset of those codes, but the key point remains: extra bits reduce the quantisation step and increase transform precision.
When a smooth gradient is represented by steps that are too coarse, neighbouring pixels cluster into visible bands. Log/PQ curves allocate code values efficiently, but heavy grading, noise reduction, keys and repeated transforms expose 8-bit limitations sooner. Dither/noise can hide boundaries but does not restore lost source precision.
Dynamic range is determined by capture/display capabilities and transfer/encoding. A 10-bit Rec.709 SDR signal can have less luminance range than an 8-bit log capture while still being quantised more finely inside its coded range. Bit depth controls granularity, not primary locations or scene latitude.
Every nonlinear transform redistributes distances in code space. Lift/curves, secondary keys, gamut transforms and HDR output can stretch a small source interval into a large visible interval. Higher source precision and internal floating-point precision reduce the risk that hidden quantisation becomes visible contouring.
In 4:4:4, luma and both chroma components have the same spatial sampling grid. This is useful for keying, compositing, text/graphics and repeated processing. But 4:4:4 says nothing about bit depth or compression: 8-bit 4:4:4 is still 8-bit, and compressed 4:4:4 can still have codec artefacts.
For every two horizontal luma samples there is one Cb and one Cr sample; vertical chroma resolution is retained. This is a common production/broadcast compromise: much lower bandwidth than 4:4:4 while handling coloured edges substantially better than typical 4:2:0 distribution footage.
In a progressive interpretation, the chroma grid has half the luma width and half the height, roughly one Cb/Cr pair per 2×2 block of luma samples. This is efficient for delivery because vision is less sensitive to fine chroma. But one-pixel coloured text, screen capture, green-screen hair and saturated edges suffer much more.
4:2:0 describes density, but chroma samples can be vertically/horizontally co-sited or centred in different ways. Codec metadata and standard conventions define the phase. Good up/downsampling is therefore not merely bilinear resizing of two smaller images; it is filter- and phase-aware resampling.
Every 4:4:4→4:2:0→4:4:4 cycle low-pass filters and resamples chroma again. After several transcodes, fine coloured edges soften and key boundaries get dirtier even at high bitrate. Intermediate/mastering codecs therefore often use 4:2:2/4:4:4 not for viewing itself but for later processing.
For gradients/log/HDR and aggressive grading, bit depth is usually more critical. For chroma key, screen graphics and fine coloured edges, chroma sampling matters more. There is no universal ranking. Ask which information the next operation needs more: tonal precision or spatial chroma detail.
Resolve processes images at high internal precision, which prevents additional arithmetic loss, but existing 8-bit banding or missing 4:2:0 chroma detail is not automatically restored. Chroma NR/upsampling/AI tools can reconstruct appearance, but that is estimation, not recovered original samples.
YouTube/OTT final delivery often requires 4:2:0, while VFX round trips, keying, archive mezzanines or broadcaster handoff may require 4:2:2/4:4:4 and 10/12-bit. Exporting 4:4:4 from a 4:2:0 source can provide a better container for further processing, but it does not invent new chroma detail.
Keep two questions separate: `how many levels can each sample take?` is bit depth; `how many chroma samples exist spatially?` is subsampling. Codec specifications become much less mysterious once those are separated.
Technical definitions are anchored to standards bodies and primary specifications wherever that distinction matters.
4:2:0/4:2:2/4:4:4 SubWidth/SubHeight and chroma-location semantics.
8/10-bit-scalable narrow-range quantisation equations.
10/12-bit signal representation and 4:2:0/4:2:2/4:4:4 formats.
Full/SDI/narrow digital mappings for 10–16-bit PQ values.
