Video Codec Engine

Nishi vs the field — every Nishi cell is measured against real organ source at emit time; each gap names the watch contract that will close it.

How the Nishi compression engine measures up -- Nishi vs x264 and x265 and SVT-AV1 and DCVC-FM (neural)

Layer 1 · Executive

Where we are. Measured 2026-08-19. A real H.264-class sovereign codec is live in two measured profiles with range coding, quarter-pel motion, RD-selected partitions, in-loop deblock, real-bit RDO over skip and transform, and structural exceeds (integer-deterministic bit-exact everywhere, one codec on three surfaces). The gap to the majors is second-generation architecture (modes and partitions RDO, larger transforms, B-pyramid) and the learned loop (a post-filter that fires, hyperprior, an end-to-end neural codec, perceptual coding). The live resolution ceiling is tracked on /compare/video.

Where we need to go. Close the second-generation architecture gap with every rung measured on the MSU set against the 822 baseline and never negative on any sequence, then the learned loop in integer inference, then the neural frontier -- keeping bit-exact determinism as the exceed the float codecs cannot match.

The unit. 1 u = one measured session-leg. Calibration from landed rungs: the mangagen panel compositor went from existing substrate to shipped and live-verified in ONE leg (2026-08-13); the citations rung went from 3 to 55 domains in one leg across seven seats (2026-08-18); a greenfield engine with a bite-proven gate has measured 2 to 4 legs. Estimates recalibrate as rungs land and PR7 actuals write back.

Cost to second-generation architecture: 6.5 u. Through M0.

Cost to the learned loop: 10.5 u. Through M1.

Cost to the neural frontier: 18.5 u. Everything below.

12 of 21 capabilities measured|3 of them measured exceeds|9 open|coverage 571/1000|adoption 6 full / 6 partial

Layer 2 · Roadmap

Do this next — computed by the ranker, never chosen by a seat

Order from nx_compare_rank (nx_dr_ocm: (deficit + cost-of-delay + option + enables) x sponsor x self-sufficiency x momentum / cost). FINISH rows are rungs whose symbol is present but whose organ is short of full adoption: the cheapest closures on this board, listed before any new work. Stamp: # asof=1787883505 domain=videocodec target_version=0.1 rungs=7 done=0 open=7 finish=0 ranker=nx_dr_ocm

#	Stage	Rung	Priority	Derivation
#1	0.1	Full-mode RDO (R0) vc_rdo_modes	1133	v=17 m=1 c=15
#2	0.1	B-frames and hierarchical GOP (R2) vb_encode	400	v=12 m=1 c=30
#3	0.1	Larger transforms (R1) vt_dct16	300	v=6 m=1 c=20
#4	later	Post-filter trained to fire (R3) vc_nf_train	266	v=4 m=1 c=15
#5	later	Perceptual and generative coding (R6) pc_gen	175	v=7 m=1 c=40
#6	later	Learned entropy hyperprior (R4) ne_model	160	v=4 m=1 c=25
#7	later	End-to-end neural codec (R5) nc_codec	75	v=3 m=1 c=40

Critical path — contract, done-rule, executor, cost

Rung	Closes with	Definition of done (pre-declared)	Executor	Est.
Full-mode RDO (R0)	vc_rdo_modes	Intra modes and inter partitions trial-coded through the real range-coder cost like skip and transform already are; BD-rate on the MSU four-sequence set vs the 822 baseline never negative on any sequence; encode fps above the RTC tier floor	Organ	1.5 u
Larger transforms (R1)	vt_dct16	16x16 and 32x32 integer transforms with RD selection beside 4x4 and 8x8; bit-exact decode; BD-rate measured on flat content where the gain lives	Organ	2 u
B-frames and hierarchical GOP (R2) after R0	vb_encode	Bi-prediction with a B-pyramid under the existing reference model; decoder bit-exact; BD-rate gain measured per sequence; latency profile kept out of the RTC default by data	Organ	3 u
Post-filter trained to fire (R3)	vc_nf_train	A training run over banked frames produces a table that measurably fires; BD-rate improves on the set; a table that changes nothing is refused, not shipped	Local model	1.5 u
Learned entropy hyperprior (R4) after R3	ne_model	A trained model over learned features (not DCT coefficients, which the measurement already showed near-optimal) with integer inference; bits per pixel vs the hand context coder published	Local model	2.5 u
End-to-end neural codec (R5) after R4	nc_codec	A DCVC-class learned codec in integer inference running on the wasm path; BD-rate vs our classical codec and vs x265 on the set; speed axis published beside it	Local model	4 u
Perceptual and generative coding (R6) after R5	pc_gen	Generative reconstruction at very low rate judged by the no-reference critic and a full-reference metric; a recon that hallucinates content fails the faithfulness tooth	Local model	4 u

Milestones

Milestone	Rungs	Cumulative
M0 · Second-generation architecture	R0,R1,R2	6.5 u
M1 · Learned loop	R3,R4	10.5 u
M2 · Neural frontier	R5,R6	18.5 u

Layer 3 · Engineering

How this is scored. Every Nishi mark is measured: the generator reads the real organ source on disk and requires the implementing symbol to exist (no self-grading). A watching tag names the organ and symbol contracted to close a gap — the mark flips itself on the next compare beat when that workstream ships, and the comparewatch- plane row flips with it. The flip is necessary, not sufficient: it proves the symbol exists, never that the capability is good. The bar is the rung's pre-declared done-rule, proven by its gate — a symbol shipped without the behaviour behind it is a defect, and the flip is exactly what makes that defect visible instead of quiet. Competitor marks record documented capability presence — presence, not depth or scale. Adoption is measured too: every measured row carries where its organ stands on the estate's ladder (source → built → promoted → registered → invoked; libraries by importer reach minus validation importers; gates by the execution surfaces that run them). A row is fully adopted only at the top of its ladder; anything short is tagged partial with the exact remedy, so a build nobody promoted can no longer read as shipped. Census stamps: importers asof 1787849099, gate census asof 1787855507 (unix seconds; -1 = census absent).

Capability matrix — measured against source

◉ leads / measured exceed● present◐ partial○ absent · click any capability for its evidence

Capability	Nishi	x264 (H.264)	x265 (HEVC)	SVT-AV1	DCVC-FM (neural)
Block transform -- integer DCT Measured: `nx_dct8_forward_1d` exists in `runtime/nx_dct8.nx`, verified at emit. Nishi 4x4 + 8x8 [h264]; HEVC to 32x32, AV1 to 64x64 (big win on flat content) Adoption: `LIB-WIRED importers=27 nonval=14` — fully adopted (top of its ladder).	●	●	●	●	●
Context/arithmetic entropy coder Open — no implementing organ is measured for this axis yet. Nishi binary range coder saves 33pct vs its own CAVLC (MEASURED); CABAC [x264] + learned entropy are deeper	○	◉	◉	◉	◉
Quarter-pel motion compensation Measured: `qp_pixel` exists in `runtime/nx_vsubpel.nx`, verified at emit. Nishi 1/4-pel translational; AV1 adds 1/8-pel, affine, OBMC [av1spec] Adoption: `LIB-WIRED importers=5 nonval=3` — fully adopted (top of its ladder).	●	●	●	◉	●
Variable block-size motion partitions Measured: `vc_enc_mb_part8_inter` exists in `runtime/nx_vcodec.nx`, verified at emit. Nishi P_8x8 = 4 independent 8x8 MVs per MB, RD-selected vs 16x16, differential MV -- LIVE vcv-7 (BUILD 810). MEASURED +2-4pct bit-exact on MSU akiyo/foreman/bus/mobile at qp8/20/32. The first partition rung; x264 splits to 4x4, HEVC uses a CTU quadtree, AV1 to 4x4-plus. Cuts the dominant P-frame residual (measured 87-97pct of P bits) Adoption: `BUILT-UNPROMOTED` — PARTIAL: compiled, never promoted to the serving root: /api/promote it. adoption BUILT-UNPROMOTED	●	●	◉	◉	●
Rich intra prediction Measured: `vc_pred_rich` exists in `runtime/nx_vcodec.nx`, verified at emit. Nishi 9 modes + MPM; HEVC 35, AV1 56-plus directional Adoption: `BUILT-UNPROMOTED` — PARTIAL: compiled, never promoted to the serving root: /api/promote it. adoption BUILT-UNPROMOTED	●	●	◉	◉	●
In-loop filtering Measured: `vv_deblock` exists in `runtime/nx_video_codec_wasm.nx`, verified at emit. Nishi deblock, gentle-tuned vcv-8 (c1 qp1.5+2 -> qp/2+2 was the FIFTH mis-tuned constant: over-smoothing the in-loop reference cost 24-47 BD-rate pts, MEASURED + shipped BUILD 815); HEVC adds SAO, AV1 adds CDEF + loop restoration Adoption:* `LIB-WIRED importers=62 nonval=17` — fully adopted (top of its ladder).	●	●	◉	◉	●
Rate-distortion mode decision (real-bit RDO) Measured: `vc_enc_block_packed_e` exists in `runtime/nx_vcodec.nx`, verified at emit. Nishi RDO now covers THREE decisions: (1) RD-SKIP (BUILD 816, archival profile): every P-MB trial-coded skip-vs-code through the LIVE range coder (est/probs snapshot+restore), argmin J=32SSE+qp^2bits -- MEASURED -7..-44 BD-rate pts; (2) per-MB HEAT-AQ skip policy (BUILD 819, live): past-streak-modulated thresholds, -1.72pct avg; (3) per-MB 4x4-vs-8x8 TRANSFORM trial (BUILD 820, live): the 803 machinery awakened -- both transforms trial-quantized per coded MB, argmin J commits, t8 bit explicit on the wire (stock decode) -- MEASURED -0.52pct avg incremental (never negative, +9.3pct encode at 120fps). Its bit cost is still the CAVLC table (the true stream is range-coded) -- replacing it with real-rc-bit trial costing is the remaining P3 rung. x264/x265 apply full-RDO to every mode+partition+transform (the 20-year integration edge); ours now covers skip+transform -- modes/partitions remain Adoption: `BUILT-UNPROMOTED` — PARTIAL: compiled, never promoted to the serving root: /api/promote it. adoption BUILT-UNPROMOTED	●	◉	◉	◉	◉
B-frames / hierarchical GOP Open — watching `runtime/nx_vcodec_bframe.nx` : `vb_encode`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. Nishi is I+P only; hierarchical-B is 15-30pct of the majors efficiency watching vb_encode	○	◉	◉	◉	●
Larger transforms (16/32/64) Open — watching `runtime/nx_vcodec_bigtx.nx` : `vt_dct16`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. Nishi 8x8 max; HEVC 32x32 [h265] [x265], AV1 64x64 watching vt_dct16	○	○	◉	◉	○
Trained restoration post-filter Open — no implementing organ is measured for this axis yet. Nishi has a trained integer post-filter but the table is mostly-identity today (barely fires); neural codecs learn the whole loop	○	○	○	○	●
Learned entropy model (hyperprior) Open — watching `runtime/nx_neural_entropy.nx` : `ne_model`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. MEASURED: our hand-context coder is already 0.088 bits/sig-flag (near-optimal) -- a trained model on DCT coeffs is marginal; the DCVC win is a JOINTLY-TRAINED model over LEARNED features, not DCT [dcvcfm24] watching ne_model	○	○	○	○	◉
End-to-end neural codec Open — watching `runtime/nx_neural_codec.nx` : `nc_codec`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. DCVC-FM beats VVC by 20-40pct [dcvcfm24]; Nishi is classical block-based only watching nc_codec	○	○	○	○	◉
Perceptual / generative coding Open — watching `runtime/nx_perceptual_codec.nx` : `pc_gen`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. The number-one frontier (momentum 108); GAN/diffusion codecs -- Nishi absent watching pc_gen	○	○	○	◐	◉
Compression efficiency (external BD-rate, ffmpeg oracle) Measured: `vc_enc_frame_packed_rc` exists in `runtime/nx_vcodec.nx`, verified at emit. MEASURED on 4 standard CIF seqs vs x264/x265/AV1 (ffmpeg Y-PSNR, GOP48 zerolatency RTC) [svtav1] [bjontegaard01] [msu-codec-comparison]. TWO PROFILES (a codec is a rate/quality/SPEED triangle -- BD-rate alone is a partial grade): (1) RTC/live (BUILD 822, the shipped /video default): THREE gen-2 rungs shipped 2026-07-12 -- P2 HEAT-AQ (819: protect long-static MBs from drift, per-MB skip-thresh /2 on a 4-skip streak; -1.72pct avg) + t8 RD-auto (820: per-MB 4x4-vs-8x8 transform trial; -0.52pct) + P3 REAL-BIT RDO (822: the trial made byte-faithful to the emit paths -- RDOQ levels in BOTH arms + each arm costed by a read-only walk of the coder's OWN adaptive sig-map bins (rc_sig_cost_q8/rc_bits_q8) instead of CAVLC tables; -2.58pct avg incremental, biggest on motion content bus -3.16) = COMBINED MEASURED -4.82pct avg BD-rate vs the 817 baseline (akiyo -6.86 / foreman -6.00 / bus -4.05 / mobile -2.38), never negative on any seq, same-scorer x264 gap cut +176pct -> +159pct in one day, encode 119 fps (measured, far above every tier floor; the real-bit cost walk is FREE vs the tables); the CHURN direction (cheapening active MBs) measured +6.4pct = REFUTED, never do it. (2) ARCHIVAL/offline (RD-skip emode bit11, ~10 fps, for stored-media transcode): avg +140pct vs x264 (old-scorer scale) -- RD-skip's J=32D+q^2R trial coding buys -7..-44 BD-pts but costs 9-19x encode, so it is NOT the live default. Lineage 811-815 fixed five mis-tuned constants (foreman 21x -> 2.4x, 8.7x bitrate cut from the original +703..+1987pct). INTRA-only +41..+66pct. Gen-1 levers measured out: subpel/partition/ME/MV-pred ~0, multi-ref 1.5pct, intra-in-P wash, SATD 0.1-0.5pct, I16-pred/16x16-TX/CDEF/QP-cascade REFUTED. Remaining vs x264 = 2nd-gen architecture (32x32 CTU, hier-refs, B-pyramid) + compound RDO Adoption: `BUILT-UNPROMOTED` — PARTIAL: compiled, never promoted to the serving root: /api/promote it. adoption BUILT-UNPROMOTED	●	◉	◉	◉	◉
Real-time encode throughput (fps budget) Measured: `vv_enc_rct8` exists in `runtime/nx_video_codec_wasm.nx`, verified at emit. MEASURED (vc_fps_bench, wasm VM, 320x256 P-frame, talking-head): Nishi RTC profile ~89 fps single-thread (browser-viable for live calls); ARCHIVAL/RD-skip profile 4.6-9.7 fps (offline only). x264 zerolatency = the RTC gold standard (1080p realtime on a desktop core) [x264]; x265/AV1 need fast presets or HW; DCVC-neural = GPU-only. ★A codec claim is INCOMPLETE without this axis -- BUILD 816 shipped an RD-skip that collapsed live encode 89->4.6 fps because the census had NO speed axis; 817 split the profiles + added this row Adoption: `LIB-WIRED importers=62 nonval=17` — fully adopted (top of its ladder).	●	◉	◉	●	○
Live resolution ceiling + per-device adaptive send Measured: `vv_enc_rct8` exists in `runtime/nx_video_codec_wasm.nx`, verified at emit. MEASURED (vc_res_bench/vc_hd_bench, wasm VM): the codec itself is RESOLUTION-AGNOSTIC (bit-exact GREEN proven at 640x480/960x544/1280x720); the LIVE ceiling is the client memory layout + single-thread fps. BUILD 818 (2026-07-12) raised the shipped ceiling 320x256 -> 640x480 with Jitsi-class per-device adaptive tiers (measured: 640x480=38.6fps / 480x384=58 / 416x320=70 single-thread; tier picked by mount-time measured encoder class; LO 128x96 simulcast for weak receivers; old-peer rooms fail safe to 320x256) + adaptive tile rendering (the old client DOWNSCALED every received frame onto a 320x256 canvas -- resolution was capped at DISPLAY too). Majors do 720p-1080p live (x264 1080p realtime single-core; Zoom/Meet/Jitsi default 720p) -- Nishi is BEHIND on ceiling until the tile-parallel HD arc lands: 720p single-thread = 8.8 fps MEASURED, needs ~4-band parallel encode with the rich codec in the band path (band path is legacy-entropy today; legacy band at 1280x704 = 251ms/band REFUTED). ★This axis was MISSING from the census while users saw potato -- presence of a speed row did not catch a resolution cap; grade what users SEE. UPDATE 2026-08-19: V1 SHIPPED (build 862) -- the CLIENT ceiling was RAISED from 640x480 to HD 1152x640/1024x576 adaptive via vc_res_tier_hd + a fully DERIVED region layout (no magic hex, grown to 108 MB, gated on all-peers-862 + deviceMemory>=4); codec bit-exact roundtrip at 1152x640/1024x576/640x480 gate-proven (25/25) and served byte-identical to the gated build. 1280x720 (45 MB-rows, no even band split) is the next K=2 rung; see V1 on /compare/video Adoption: `LIB-WIRED importers=62 nonval=17` — fully adopted (top of its ladder).	●	◉	◉	◉	●
Sovereign own-stack (zero third-party) Measured exceed: `vc_t8_init` in `runtime/nx_vcodec.nx`, verified at emit. Own language, compiler, wasm -- x264/x265/AV1 are C/C++, DCVC is Python/CUDA Adoption: `BUILT-UNPROMOTED` — PARTIAL: compiled, never promoted to the serving root: /api/promote it. adoption BUILT-UNPROMOTED	◉	○	○	○	○
Integer-deterministic bit-exact everywhere Measured exceed: `vc_mb_t8_ok` in `runtime/nx_vcodec.nx`, verified at emit. No-float pipeline reproduces bit-exact on ANY device, which is what makes recon-hash conformance possible; float codecs cannot Adoption: `BUILT-UNPROMOTED` — PARTIAL: compiled, never promoted to the serving root: /api/promote it. adoption BUILT-UNPROMOTED	◉	○	○	○	○
One codec
browser tab + Nishi browser + native NishiOS Measured exceed: `vv_enc_rct8` in `runtime/nx_video_codec_wasm.nx`, verified at emit. The same sovereign codec runs in a tab, the Nishi browser, and native NishiOS; others need hardware or wasm ports Adoption: `LIB-WIRED importers=62 nonval=17` — fully adopted (top of its ladder).	◉	◐	◐	◐	○
Full-mode RDO over intra modes and partitions Open — watching `runtime/nx_vcodec.nx` : `vc_rdo_modes`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. The RDO row's own remainder: skip and transform are RD-decided, modes and partitions are not watching vc_rdo_modes	○	◉	◉	◉	◉
Restoration post-filter trained to fire Open — watching `runtime/nx_vcodec_nf.nx` : `vc_nf_train`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. The trained table is mostly identity today; a training run that measurably fires, BD-rate proven watching vc_nf_train	○	○	○	○	●

On these two registers. Rows are declared in the domain's plan file and carry the debt id, which is the join key back to the sovereign debt plane — that plane, not this page, is the authority on state. Reconciling them automatically (the regen reading the plane and refreshing these rows) is a named, owed rung; until it lands, treat an id here as a pointer to look up, not a status to trust.

Honest verdict. Nishi runs a REAL H.264-class sovereign codec end-to-end -- range-coded entropy, quarter-pel motion, RD-selected partitions, in-loop deblock, live at nishifamily.com/video in two MEASURED profiles (RTC ~89fps at +159pct BD-rate vs x264; ARCHIVAL RD-skip at +140pct) with a 640x480 adaptive live ceiling (BUILD 818). It is not yet a bitrate rival: the remaining gap vs the majors is second-generation architecture (32x32-plus CTUs, hierarchical B-refs, full-mode RDO) and the live ceiling trails their 720p-1080p until the tile-parallel HD arc lands. The genuine exceeds are structural -- the whole stack in one sovereign language, integer-deterministic bit-exact reconstruction on any device (recon-hash conformance the float codecs cannot do), and one codec running browser tab, Nishi browser, and native NishiOS. Every number on this page is MSU-methodology measured (ffmpeg Y-PSNR oracles), liar-killed, and re-graded on every ship -- including the regressions.

Person · product · place — not yet measured for this domain

Every compare carries this layer. Declare knowledge/compare/videocodec.ppp (rows surface|nishi or c1..c4|label|url|connect naming OUR live surface and each rival's front door), run nx_ppp_probe domain videocodec, and this section fills itself on the next beat: the same ruler on both sides — privacy and CX (third-party hosts, tracker classes, cookies, security headers), design and longevity (design hygiene, computed WCAG contrast, render-blocking resources, unsized media, script weight, theme and motion queries), findability (landmarks, skip link, on-site search, breadcrumb, headings, internal links).

References

Beyond a link list. Every reference below resolves twice — the publisher's copy and, where banked, the estate's own non-rottable library mirror with a content pin — and carries its evidence class plus the exact claim on this page it grounds. Keyed marks like [key] in the matrix notes jump here. A dash means honestly absent, never assumed.

[x264] VideoLAN. x264, the best H.264/AVC encoder: CAVLC/CABAC entropy coding, adaptive B-frame placement and B-frames as references, psy-RD and psy-trellis, multipass ABR/VBV ratecontrol, 4 or more 1080p streams in realtime on a consumer machine. publisher · read in our library knowledge/fetched/cmp_videocodec_x264.html · pin h8ac7051331232bc3423f7c294ee026b062290eb8459679751e822e5f44894a79 · accessed 2026-08-18 · vendor-docGrounds: The x264 column: "Context/arithmetic entropy coder" (CABAC), "B-frames / hierarchical GOP" (adaptive B placement, B-refs), "Rate-distortion mode decision (real-bit RDO)" (psy-RD/trellis over every mode) and "Real-time encode throughput (fps budget)" -- the page's 1080p-realtime claim is the RTC gold standard the row grades our 89 fps 320x256 against.
[x265] VideoLAN and MulticoreWare. x265, free H.265/HEVC encoder: quad-tree prediction, adaptive B-frame placement, CABAC, multiple reference frames, frame-level and wavefront parallelism; GPL with commercial licence. publisher · read in our library knowledge/fetched/cmp_videocodec_x265.html · pin hb0c55e0005ad2b44ece4b190fc6151f17f37eb96afe8f81089bd77537f176fac · accessed 2026-08-18 · vendor-docGrounds: The x265 column: "Larger transforms (16/32/64)" and "Variable block-size motion partitions" ride the HEVC CTU quadtree the page names; "Real-time encode throughput" code 2 with fast presets or parallelism, as the page describes.
[svtav1] Alliance for Open Media. SVT-AV1 -- the AOM-maintained AV1 software encoder (Clear BSD; official GitLab repository with README, CHANGELOG and releases). publisher · read in our library knowledge/fetched/cmp_videocodec_svtav1.html · pin h29f5ebda4e8ba42867599b0979acccdc377d006f15d12c732b3cc90e4712a7e5 · accessed 2026-08-18 · vendor-docGrounds: The SVT-AV1 column: the AV1 encoder our ffmpeg oracle runs for the "Compression efficiency (external BD-rate, ffmpeg oracle)" row and whose fast presets ground the "Real-time encode throughput" code 1.
[av1spec] Alliance for Open Media. AV1 Bitstream and Decoding Process Specification (v1.0.0 with errata), the normative decoder definition: 1/8-pel and warped/affine motion, OBMC, directional intra modes, CDEF and loop restoration, transforms to 64x64. publisher · read in our library knowledge/fetched/cmp_videocodec_av1spec.html · pin hacf58a3fdb82a85e25aa7fe19b8fe2a5ccc02884d0e852f2fc9b05ba7aa49609 · accessed 2026-08-18 · published-standardGrounds: The AV1 feature reads across "Quarter-pel motion compensation" (AV1 adds 1/8-pel, affine, OBMC), "Rich intra prediction" (directional modes), "In-loop filtering" (CDEF + loop restoration) and "Larger transforms (16/32/64)" (AV1 64x64) are read from this normative text, not from marketing.
[h264] ITU-T. Recommendation H.264: Advanced video coding for generic audiovisual services (first edition 05/2003; in-force editions through 2024) -- ITU-T publication page listing every edition. publisher · read in our library knowledge/fetched/cmp_videocodec_h264.html · pin hce4404ca022f3b50d0ba54439285cac85dcbc09d0a6ec6a6059642f833b1ba84 · accessed 2026-08-18 · published-standardGrounds: The @verdict claim "a REAL H.264-class sovereign codec": CAVLC/CABAC, quarter-pel MC, 4x4/8x8 integer transform, 9 intra 4x4 modes and in-loop deblocking are H.264 toolset elements, so "Block transform -- integer DCT", "Rich intra prediction" (Nishi 9 modes + MPM) and "In-loop filtering" (deblock) are graded against this standard's toolset.
[h265] ITU-T. Recommendation H.265: High efficiency video coding (first edition 04/2013; in-force editions through 2024) -- ITU-T publication page listing every edition. publisher · read in our library knowledge/fetched/cmp_videocodec_h265.html · pin h0f621ca591bf5c68bdb52451098015cb033d6aba38178de88422da585eec886b · accessed 2026-08-18 · published-standardGrounds: The HEVC feature reads: "Larger transforms (16/32/64)" (HEVC 32x32), "Rich intra prediction" (HEVC 35 modes), "In-loop filtering" (HEVC adds SAO) and the CTU quadtree behind "Variable block-size motion partitions" are HEVC toolset facts this standard defines.
[dcvcfm24] Li, Li, Lu. Neural Video Compression with Feature Modulation (DCVC-FM). CVPR 2024, arXiv:2402.17414 -- abstract headline: 29.7 percent bitrate saving over the previous SOTA neural video codec under the single-intra-frame setting with 16 percent fewer MACs. publisher · read in our library knowledge/fetched/cmp_videocodec_dcvcfm24.html · pin ha127e0bee1abebea0a55a8339570014a261eff241357000780bdb40d3fe3cf02 · accessed 2026-08-18 · published-paperGrounds: The DCVC-FM (neural) column: "End-to-end neural codec" and "Learned entropy model (hyperprior)" code 2 are this paper's architecture (jointly trained model over learned features, which the entropy row's note names as the real win). The row note's "beats VVC by 20-40pct" is a read of the paper's VTM comparison tables, not of the abstract, whose stated number is the 29.7 percent above.
[bjontegaard01] Bjontegaard, G. Calculation of average PSNR differences between RD-curves. ITU-T SG16 Q.6 VCEG document VCEG-M33, Austin TX, April 2001 (the BD-PSNR / BD-rate definition). publisher · read in our library knowledge/fetched/cmp_videocodec_bjontegaard01.doc · pin h37cfd46e88a3e8b880c95bc92753027c8fed5fde39947b0a50b17efc0e336014 · accessed 2026-08-18 · published-standardGrounds: "Compression efficiency (external BD-rate, ffmpeg oracle)": every +/- pct on this page (+159pct RTC, +140pct archival, -4.82pct avg vs the 817 baseline) is a Bjontegaard delta-rate over four rate points as this document defines; the mirror is the original ITU-hosted Word document.
[msu-codec-comparison] Moscow State University Graphics and Media Lab. MSU Video Codecs Comparison (annual reports 2003-2025): objective Y-PSNR/SSIM/VMAF and subjective methodology over standard sequences at CIF through 4K, the reference public codec benchmark. publisher · read in our library knowledge/fetched/cmp_videocodec_msu.html · pin hf38da4818d416b566c54ccf26e07b3f66240dc00b81ec7a0d3b56b32052dbbc1 · accessed 2026-08-18 · published-paperGrounds: The @verdict claim "Every number on this page is MSU-methodology measured (ffmpeg Y-PSNR oracles)" and the akiyo/foreman/bus/mobile CIF sequence set used by nx_vcodec_rd_bench: this is the methodology page those measurements follow.

generated by nx_swcompare_matrix (sovereign NishiLang organ) from knowledge/compare/videocodec.matrix · every Nishi cell verified against organ source at emit time · watch cells re-measured on every compare beat · zero JS, zero trackers