4K Nature — 1/4.4 of the original.
Same 20-second 4K30 nature timelapse. The right side is NormH.264 at 25 MB, down from 110 MB. Both auto-play below — judge for yourself.
Heavy demo (~135 MB total) — on metered connections, click each video's play button instead of relying on auto-play.
Source — H.264 4K30 · 110 MB · ~45 Mbps
4k_nature_h264_source.mp4 · 20.35 s · high-bitrate H.264 master
4k_nature_h264_source.mp4 · 20.35 s · high-bitrate H.264 master
NormH.264 — H.264 4K30 · 25 MB (−77.3%, 1/4.4)
4k_nature_norm_h264.mp4 · VMAF 80.47 / VMAF min 68.90 · same H.264, plays everywhere
4k_nature_norm_h264.mp4 · VMAF 80.47 / VMAF min 68.90 · same H.264, plays everywhere
110 MB → 25 MB
Same H.264 codec, same 4K30, plays unmodified on every device. Judge it with your own eyes — the two clips above are the evidence.
An honest caveat about this clip — what it does and does not demonstrate.
The −77.3% above is measured against a 45 Mbps master — a source with plenty of slack in it. What it demonstrates is that a bloated input can be brought down to a sane bitrate without visible damage. It is not a proof of codec efficiency.
When we put the same 4K nature material against plain x264 at equal size (about 8 MB), x264 scored VMAF 70 and NormH.264 scored VMAF 69 — plain x264 came out slightly ahead. On this clip there is no NormH.264-specific win at equal quality.
The material where we do win at equal quality is what the g2 check flags as a fit, and that measurement is −4–5% — see the full ledger, wins and losses →
The −77.3% above is measured against a 45 Mbps master — a source with plenty of slack in it. What it demonstrates is that a bloated input can be brought down to a sane bitrate without visible damage. It is not a proof of codec efficiency.
When we put the same 4K nature material against plain x264 at equal size (about 8 MB), x264 scored VMAF 70 and NormH.264 scored VMAF 69 — plain x264 came out slightly ahead. On this clip there is no NormH.264-specific win at equal quality.
The material where we do win at equal quality is what the g2 check flags as a fit, and that measurement is −4–5% — see the full ledger, wins and losses →
Notice the playback itself. The source on the left (110 MB) buffers, stalls, hesitates on most connections. NormH.264 on the right (25 MB) plays smoothly the moment you load the page. The compression delta is also a delivery and UX delta — viewers do not need to know about VMAF; they feel the difference as “starts faster, never stutters”.
Why this is hard
4K nature timelapses are texture-heavy: leaves, grass, water, sky gradients, sun rays. There is no flat region for an encoder to give up cheaply. Plain H.264 at this quality routinely sits at 40–50 Mbps. NormMAP's commutator-norm hint tells libx264 where the eye actually pays attention, so bit budget collapses on perceptually-quiet blocks while preserving the busy ones.
Spec
| Source | 4K-14690379_3840_2160_30fps.mp4 (open-web 4K nature timelapse) |
| Resolution | 3840 × 2160 (4K UHD), 30 fps, 20.35 s, 610 frames |
| Source codec | H.264 (yuv420p), bitrate ~45 Mbps |
| Output codec | H.264 (yuv420p), via NormH.264 v0.3.2 (slime-2 preset) |
| NormMAP config | commutator-norm block-level bit allocation, libx264 ROI side-data path, aggressive bias |
| VMAF mean | 80.47 (measured on this output. VMAF tends to understate NormMAP, so please judge by eye) |
| VMAF min | 68.90 (worst frame on a heavy-detail moment) |
| Size delta | −77.3% (1/4.4) versus the 4K H.264 source (45 Mbps). Against-source figure — depends on the input bitrate |
| Decode compatibility | H.264 — every browser, device, set-top, including Safari iOS without polyfills |
See more: CityHall 4K compare (1/2 from already-AV1) · 30-day trial →