NormH.264 — 4K to 1/4 size, visually identical

NormMAP commutator-norm driven H.264 encoder. Empirical on a real 4K source: 110 MB → 28 MB (-74.3%) at VMAF mean 83.25, visually indistinguishable from the original. Drop-in FFmpeg pipeline; 100% playback compatibility (every H.264 player, hardware decoder, and edit suite works unchanged). 30-day HW-fingerprint locked trial. SlimeCodec Convert v0.3.3.

Why this matters for VOD / surveillance / archives: cut storage and CDN bandwidth to 1/4 without switching to AV1 (no new GPUs, no player upgrades, no transcoding farm overhauls). Every existing H.264 endpoint plays the output unchanged. AV1 catch-up not required.

What's new in v0.4.2 — Trial Signature Mode (default ON)

The honest history. v0.3.x trial output (releases up to v0.3.6, 2026-05-06 through 2026-05-09) carried visible monochrome rate-control artefacts in the watermark band — particularly when motion crossed the SLIMECODEC TRIAL logo area. These were not random encoder noise; they were the predictable consequence of libx264's internal AQ-mode and MB-tree heuristics treating the static high-contrast watermark as “maximum importance”, concentrating bits there at the expense of neighbouring moving content. Five days of investigation (revisions r1–r3, strength=0 bypass, 10-bit pixel-format check, diagnostics A–C, q-level CNP) failed to remove these artefacts via the conventional preserve-direction ROI path.

The fix in v0.4.2. Reverse the polarity. Instead of asking libx264 to preserve the watermark MBs, push them to maximum positive qoffset (deliberate degradation). The libx264 MB-tree no longer concentrates bits on the watermark band, no preserve-driven artefact appears, and the visible “SLIMECODEC TRIAL” logo plus the “ノイズはトライアルの制限” chroma-plane disclaimer remain unchanged. Output is bit-exactly compatible with any standard H.264 decoder.

Trial signature philosophy. The trial mark is intentional — a small visible logo + JP disclaimer + a faint mosaic in roughly the top 10 % d-value MBs that are concentrated in the watermark area. The rest of the frame, including buildings, sky, and moving subjects outside the watermark band, is rendered with the codec's full NormMAP-driven quality. This is the v0.4.2 contract: trial = visible mark, not visible damage.

Reproducing v0.3.6 behaviour. Pass --no-mosaic to revert. The artefacts will return.

Default flags in v0.4.2.

Trial Signature Mode: ON — --no-mosaic to disable.
Constrained Norm Projection (CNP): default ON in non-trial paths, automatically disabled when mosaic is on (CNP smoothing would dilute the binary mosaic field).
Mosaic coverage: 0.10 — targets the watermark area only.
Mosaic strength: 1.0 — maximum positive qoffset on the top-d MBs.

v0.5.0-trial — pure Rust in-process watermark (structural fix, single binary)

v0.4.2 trial contract stays in place, and v0.5.0 isolates the watermark from the encoder's decision path entirely. The "SLIMECODEC TRIAL" text is drawn by the binary itself, using a pure-Rust 8x8 bitmap font, into the chroma planes (U/V) only. The Y plane is never touched — and since libx264's mb-tree / AQ / motion estimation and NormMAP all read Y exclusively, the watermark is structurally invisible to every encoder decision system. No subprocess, no ffmpeg.exe to ship, no external dependency for the overlay — single binary.

Consequences:

The watermark-area rate-distortion interference of v0.4.x — which originated from the encoder treating the overlay as a special case — is structurally eliminated. NormMAP runs commutator-norm analysis uniformly over the whole frame; no ROI exception for the watermark area.
Pass 1 (NormMAP analysis) and Pass 2 (encode) draw the same watermark, so analysis and encoder see identical texture.
The production (paid) build skips the watermark draw entirely — same encoder.
Linux: only the libav* shared libraries are required (the ffmpeg binary is not). Windows: binary + bundled FFmpeg DLLs ship in a single zip; unzip and run.

Trial v0.5.0 vs Production differences

Aspect	Trial v0.5.0	Production (paid)
Pipeline	1 stage (in-process watermark + SlimeCodec encode)	1 stage (no watermark)
Watermark	"SLIMECODEC TRIAL" + subscribe URL drawn into chroma (U/V) planes only, bottom-center (Y untouched, invisible to all encoder decision systems)	none
Processing time	baseline + ~0.1% (Y-plane draw is ~1 ms / 4K frame)	baseline
Temporary disk	none	none
External dependency	Linux: `libav` system libs / Windows: bundled DLLs only. No ffmpeg binary required.*	same
NormMAP compression	unchanged (NormMAP core untouched)	unchanged
Watermark artefact	none — Pass 1/Pass 2 see identical texture and there is no ROI exception	n/a

Subscription (no watermark): see the subscription panel below ↓ for tier details (card checkout is being set up; existing trial accounts will be notified by email when it opens).

Note: this architecture resolves the v0.4.x watermark interference (5 days of investigation) by feeding identical pixels to Pass 1 and Pass 2 and removing the ROI exception for the watermark area.

Visual proof — 4K side-by-side

Same 4K nature frame, original vs NormH.264 slime-1 (1/4 size). Click any image for full resolution.

Center crop (zoom on detail to inspect):

left: original 110 MB · right: NormH.264 28 MB (-74%, VMAF 83.25)

Pricing — 6 tiers (paid)

The paid build differentiates conversion speed (thread × preset) and feature caps per tier. The throughput gap between Trial and OEM is roughly 150× — price-aligned product differentiation.

Tier	Price	Resolution	Machines / jobs	Speed (4K 30fps 1-min clip)	Use case
Free Trial	free	4K	1 / 1	~25 min (chroma watermark)	Evaluation — quality / compression check
Individual	$20 / yr	1080p	1 / 1	~6 min	Personal / hobby (no commercial bulk use)
Pro ★	$69 / mo	4K	2 / 2	~3 min	Real entry-product — small studios / YouTubers / internal PoC
Corporate	$799 / mo	4K	5 / 4	~2 min	Corp business / archive / VOD / surveillance (invoice billing OK)
Corp+ / XDCAM	$1,999 / mo +	8K + XDCAM path	10 / 8	~30 s (faster than realtime)	Broadcasters / video archive / business SDK + 24h SLA
OEM / Enterprise	individual contract	unlimited	unlimited	~10 s (6× realtime)	Cloud / embedded / streaming infra, API + white-label, annual

Conversion speed is differentiated per tier by enforced thread count × libx264 preset (e.g. Individual = 2 threads × slow, Corporate = 8 threads × medium, OEM = unlimited × fast).

Card checkout is being set up — once Stripe is wired, per-tier purchase buttons will appear here. Existing trial accounts will be notified by email when subscriptions open; if you haven't registered yet, please sign up via the form below.

OEM / Enterprise (large-scale cloud / embedded / streaming infra) is custom contract regardless of the checkout system. For deployment discussion, contact sasaki@javatel.co.jp directly (NDA available).

Download — Login / Register

30-day HW-fingerprint locked trial. Already registered? Use the Login tab — enter email + password and we email a fresh download link (valid 24 h). New users use the Register tab; you'll receive a confirmation email and clicking YES issues your link.

Credential persistence: The email + password you set here are the long-lived account credentials and will gate the production (paid) build later. One email = one account.

Email (your ID) * Password * Build *

Not registered yet? Switch to Register tab →

Name * Email (your ID) * Password * · 8+ characters Password (confirm) * Country *

Account type * Individual Corporate

Company / organisation (optional — required for Corporate) Build * I understand this is a 30-day HW-fingerprint locked trial; the binary will refuse to run after expiry.

4K empirical (real source measurement)

Tested on a 4K nature clip (3840×2160 30fps, 20.4 sec, H.264 AVC):

Output	Size	Reduction	VMAF mean	VMAF min	Visual
Original (45 Mbps)	110 MB	—	(reference)	—	—
NormH.264 slime-1 (recommended)	28 MB	-74.3%	83.25	72.31	indistinguishable
NormH.264 slime-2 (aggressive)	25 MB	-77.2%	80.47	68.90	slight loss

VMAF 83 is the threshold below which most viewers notice degradation. slime-1 lands at 83.25 mean — empirically the sweet spot where storage is quartered with no visible quality loss in real-world viewing.

What's new in v0.3.3

v0.3.3 keeps the v0.3.2 NormMAP encoder unchanged and adds three things:

Unified output codec switch: --output-codec h264|av1. h264 (default) routes through libx264 with full NormMAP ROI; av1 routes through libsvtav1 (research preview — native SVT-AV1 ROI map dispatch is queued for v0.4).
Trial-mode watermark: every trial output now carries a small Javatel logo + SLIMECODEC TRIAL banner burned into the bottom-right of every frame. License activation is still deferred to v0.4; in v0.3.3 the watermark is always on.
30-day trial hardened (offline-only): state is now mirrored to a second anchor location, the OS file mtime acts as an external forward-only time floor, and primary/secondary disagreement triggers immediate trial expiration. Closes the simple clock-rollback and single-file-delete bypasses.

The H.264 path is unchanged from v0.3.2 — same NormMAP ROI behavior, same speed, same auto-tune. Existing v0.3.2 trial state is preserved on first v0.3.3 launch.

Performance history (v0.3.1 → v0.3.2)

The H.264 encoding path in v0.3.3 inherits the v0.3.2 numbers below (unchanged), so this table also reflects current v0.3.3 H.264 behavior.

Metric	v0.3.1 stock	v0.3.2 (this build)	Change
4K Pass 1 wall time	17.8 s	6.9 s	2.6× faster ✅
4K full convert wall	27.4 s	16.6 s	1.65× faster ✅
Bitrate savings (4K surveillance, vs plain libx264)	-37.1%	-45.4%	+8.3 pp ✅
VMAF on the above	99.2	89.6	−9.6 pts (still >83) ⚠

If you need the v0.3.1 conservative behavior, set --quality-bias +0.02 (see Quality bias section below).

Speed improvements come from:

AVX2 + FMA SIMD path for commutator-norm math
12-thread per-block parallelization
Frame pipelining (decoder ↔ compute overlap)
Content-aware auto-tune (surveillance / shake / sport / nature / generic auto-detection)

Quick start (Windows GUI, recommended)

Download the GUI + CLI zip and extract to any folder
Double-click slimecodec-gui.exe
Click "Browse..." to select a video, then "Convert"
Output is created next to the input as *_slime.mp4

Quick start (Windows CLI)

Extract the zip to any folder

In cmd / PowerShell:

slimecodec-convert.exe my_video.mp4 -o output.mp4 --auto-tune --keyint 2

Or run the bundled test_encode.bat

Quick start (Linux)

tar -xzf slimecodec-convert-linux-x64-phase4_2026-05-04.tar.gz
cd slimecodec-convert-linux-x64-phase4
./slimecodec-convert input.mp4 -o output.mp4 --auto-tune --keyint 2

Quality bias dial

Fine-tune the size/quality trade-off on top of the auto-tune classifier with a single dial — slider in the GUI, --quality-bias flag in the CLI:

Setting	Effect	When to use
`--quality-bias -0.02`	Higher strength, smaller files	Compression-priority, accept slight quality drop
`--quality-bias 0.0` (default)	Balanced	Most cases
`--quality-bias +0.02`	Lower strength, larger files, higher VMAF floor	Quality-priority, distribution use

Range -0.10 .. +0.10, step 0.01 recommended. Only effective with --auto-tune.

Trial spec (Phase 5 license v2): Expires after the earlier of 30 days from first launch or 1000 invocations. Hardware-fingerprint locked (bound to a single machine), encrypted license file (~/.slimecodec/license_v2.dat / 32 bytes XOR + integrity tag). PC date rollback to extend the trial is detected and rejected.

System requirements: Windows 10/11 x64 or Linux x86_64, CPU with AVX2 + FMA support (Intel Haswell or later, AMD Excavator or later — roughly 2014+ x86 CPUs).

AV1 path

For AV1-only delivery (newer GPU / iOS 17+ / browsers with AV1 decode), see NormAV1 — research preview. Note: AV1 hardware playback is still ~50% market reach; NormH.264 is recommended for general use.

Contact

For commercial license / subscription inquiries: sasaki@javatel.co.jp

SlimeCodec is proprietary software. Patent pending JP 2026-046609.