Cross-platform video encoding suite for Termux (Android), Linux, macOS and Windows — bash + PowerShell.
v97 — 198 bugs fixed · 283+ features · ~63 950 LoC · 263 files
A smart FFmpeg wrapper for real-world video work — HDR/Dolby Vision detection and preservation, codec-aware tool dispatch (HEVC vs AV1), adaptive CRF per resolution and encoder, batch encoding with resume, action-cam telemetry extraction, and metadata-only DV/HDR10+ transforms. Built around an interactive menu that asks one targeted question per file (e.g. "this source is DV Profile 5 — preserve, convert to HDR10, or stream-copy?") instead of hand-crafting ffmpeg flags.
The same workflow runs identically across all four platforms — bash and PowerShell branches are mirrored feature-for-feature, and .conf profiles keep settings reproducible for batches and CI. The goal is to keep Dolby Vision RPU, HDR10+ dynamic metadata, 10-bit pixel formats and color signaling intact end-to-end — the kind of detail raw ffmpeg calls usually mishandle without very specific flags.
- Dolby Vision preserve end-to-end — HEVC↔AV1 cross-codec (P5/P7/P8.1↔P10), SW + all 6 HW backends (v46), HDR10+ co-existence when source carries both layers
- Dolby Vision now recognized by TVs (v70–v72) — the DV hybrids the suite builds carry the full DV layer in the bitstream but lacked the container signaling TVs decide by; when the right tool is available, the suite writes the DOVI configuration record so TVs engage Dolby Vision (PC players already read the in-stream RPU). Coverage is now complete: HEVC in MKV (
mkvmerge, v70) + HEVC in MP4/MOV (MP4Box/GPAC, v71) + AV1 in MKV (mkvmerge, v71) + AV1 in MP4/MOV (MP4Box/GPAC, v72). Optional and silent when absent (falls back to in-stream-only); applies to DV preserve at encode, HDR/DV transforms, Mux/Remux, and stream-copy/trim/audio-only passthrough. Track metadata (audio/subtitle languages, chapters) preserved faithfully. Resolves the v69 "known limitation" - HDR10+ on APV (v69) — a first: dynamic HDR10+ metadata survives encoding to Samsung APV (ffmpeg alone drops it). A dedicated stdlib engine writes the ST 2094-40 metadata per frame straight into the APV bitstream alongside static mastering-display/MaxCLL, from any HDR10+ source (HEVC / AV1 / APV) — and back: an HDR10+ APV file re-encodes to x265/AV1 with metadata intact. Output validated byte-exact against the HDR10+ ecosystem reference tools and the official test clip; optional decode-check via the OpenAPV reference decoder (
tools/openapv_validator) - HDR-aware everywhere — HDR10 · HDR10+ dynamic · DV · HLG · LOG (Apple / D-Log M / Samsung) detected automatically; per-source dialogs propose the right transform
- Spatial audio preserved & signaled — Dolby Atmos and DTS:X (v87) — object-based audio can't be created or re-encoded with free tools (the encoders are Dolby/DTS-licensed); it only survives a 1:1 track copy. The suite now detects Atmos (E-AC-3 JOC / TrueHD) and DTS:X per track, marks them in every selection dialog, and guards re-encodes with a copy-first prompt (
AV_ATMOS_POLICY/AV_DTSX_POLICYfor CI — non-interactive runs default to copy, nothing is lost silently). On MP4/MOV it also rewrites the container-level Atmos signaling that ffmpeg omits (via MP4Box, mirroring the v70-v72 DV container work) — Apple-ecosystem players decide by that box, everything else reads the bitstream. Media analysis labels the tracks, Demux marks them, Trim & Concat warns before the audio choice, and Remux now honestly drops container-incompatible spatial tracks instead of failing outright. Auro-3D (steganographic height channels in the LSBs of a normal 5.1 track) is undetectable by design — a 1:1 copy keeps it, any re-encode loses it - Eclipsa Audio (IAMF) — author it AND keep it (v88) — the open, royalty-free spatial audio format from AOMedia (pushed by Samsung/Google on recent TVs and phones). IAMF isn't a codec: it's a stream GROUP (Audio Element + Mix Presentation) over Opus substreams living in the container. The audio-only menu (option 10) now authors IAMF from stereo/5.1/7.1 sources — scalable layers (stereo base + full bed), Opus substreams, video copied 1:1 alongside; ffmpeg builds the IAMF structure natively but can't write it into a container, so MP4Box packages it (MP4/MOV only — Matroska has no IAMF mapping yet). And since ffmpeg silently FLATTENS an existing IAMF group to plain Opus tracks on any mux/copy (exit 0, nothing visibly lost — the audio twin of the v70–v72 DV container signaling), every 1:1 path (encode with audio copy, stream copy, audio-only, Remux) re-grafts the group from the source byte-identically. The audio dialog treats the substreams as one atomic unit (copy as a group, no per-track selection, no loudnorm), media analysis labels tracks (Eclipsa), and Trim & Concat warns that timeline edits flatten the group
- Atmos → Eclipsa 7.1.4, objects included (v89) — IAMF authoring now also takes 12-channel (7.1.4) sources — 7 Opus substreams (5 coupled pairs + C + LFE), scalable stereo + 7.1.4 layers — and, the bigger half: on a Dolby Atmos source (E-AC-3 JOC or TrueHD), the objects can be positionally rendered into 7.1.4 channels via Cavern (VoidXH — the only free renderer that decodes Atmos objects; TrueHD handled through truehdd, which Cavernize fetches automatically) and authored into Eclipsa with real height channels. Option 10 offers render / bed-only / skip (
AV_ATMOS_ECLIPSA_POLICYfor batch; non-interactive defaults to bed, render failure falls back to bed honestly). Video, subtitles, chapters and DJI native GPS still come from the source 1:1. Optional tool viatools/cavernize_installer(Windows validated; macOS manual guide; Linux/Termux honestly unavailable — without it everything behaves like v88). Validated on the official Dolby Tone 7.1.4: all 12 channels live in the final IAMF, heights included. DTS:X stays bed-only (no free object decoder exists). v90 closes the last Eclipsa gap: Burn-in (HUD/subtitle overlays re-encode the video but copy the audio, which silently flattens the IAMF group) now re-grafts the group onto the full MP4/MOV output too — previews (cut clips) stay honestly ungrafted, since truncated substreams can't be regrouped - DJI Osmo Action 6 D-Log M detection (v62) — Action 6 reports
bt709identically for Normal and D-Log M (the LOG curve lives only in the pixels); the real flag sits in thedjmdtelemetry track, which exiftool doesn't expose. A shared stdlib reader parses the djmd protobuf and tags D-Log M correctly so the LOG dialog (apply Rec.709 LUT / keep LOG) shows up. LOG/LUT flow also hardened: 10-bit detection via pixel-format fallback, HLG no longer misread as LOG, correct Rec.709 tagging across all containers (MP4/MOV/MKV —setparamsafterlut3d, since the filter rewrites pixels but not color metadata). v83: brand-aware LUT matching — LUTs are now recognized by their real names (AppleLog*,*Samsung*Log*,*D-LogM*), not just a fixed prefix, so the brand's LUT is surfaced first (default) while the full list stays selectable for manual choice; applies to encode, burn-in and trim/concat. Validated end-to-end on a real iPhone 17 Pro Max Apple Log clip (ProRes 10-bit → detected as Apple Log → correct Rec.709 on x265 / AV1 / burn-in / trim-concat). v85: Apple Log detection now survives trimming/remuxing — it also keys off the video stream's encoder tag (Apple ProRes), which persists through-c copy, so a clip trimmed with Trim & Concat before encoding still gets the right brand LUT (no false positives: a suite-re-encoded ProRes isn't mistaken for Apple) - DJI native GPS preserved on encode (v78) — encoding a DJI clip (Osmo Action 6 etc.) used to silently drop the native GPS telemetry track (its
codec=noneformat can't pass through ffmpeg into any container). Now the suite re-grafts the GPS track onto the MP4/MOV output after encoding (via MP4Box) — on full re-encode, fast stream-copy, and audio-only alike (anywhere the picture stays 1:1). The bulky debug track is dropped; on MKV (which can't hold it either) the suite points you to the telemetry embed option. Controlled byDJI_PRESERVE_META(auto/on/off). The old "switch to MKV to keep tracks" option — which actually crashed the encode (those tracks can't enter MKV) — is gone, replaced by this honest, automatic path - 6 HW backends, uniform UX — NVENC · QSV · VAAPI · VideoToolbox · AMF · MediaCodec with a single
1..7preset table mapped per backend - Hardware encoding correctness audit (v75) — across all 6 backends: fixed inverted quality presets on VAAPI / VideoToolbox (picking "quality" no longer gave you "fast"), AV1 HW now offered only on GPUs that can actually run it (Intel Arc / Core Ultra+, with a live capability probe on Linux), NVENC constant-quality VBR, explicit HDR10 color signaling on every card, and an honest warning that HDR10+ degrades to static HDR10 on HW encode. MediaCodec recognizes modern Snapdragon (
QTI) / Exynos (codename) SoCs and no longer claims nonexistent mobile AV1 HW encode (→ clean SW fallback). On Windows, Dolby Vision Profile 8.1 / 8.4 and AV1 DV — which carry a standard/empty codec tag rather than the DV-dedicated one — are now detected from side-data on the encode path, so the preserve dialog appears and DV is no longer silently lost; P8.4 (HLG-compatible) is also no longer mislabeled as Profile 5 in MP4/MOV - HDR10+ preserved on hardware encode (v76) — every HW encoder drops dynamic HDR10+ (only static HDR10 survives); when the HDR10+ tool is present the suite re-injects the per-scene metadata onto the HW-encoded bitstream, restoring dynamic HDR10+ intact on HEVC and AV1 — including hybrid DV + HDR10+ sources (both layers kept, DV signaled via the container record). On an HDR10+ source picked for HW, preserve-via-inject is the default. Also new: Dolby Vision Profile 7 → 8.1 in HDR/DV tools — P7 (dual-layer, typical of UHD discs) isn't read by many TVs/players, so the suite converts to universal 8.1, keeping the HDR10 base + DV metadata and warning before discarding an enhancement layer that actually carries luminance (FEL-safe;
DV_P7_FORCE=1to override). Same conversion is now applied automatically when re-encoding a P7 source with DV preserve, so the output is valid single-layer 8.1. Validated on QSV HEVC; other HW backends spec-level. v91 validates the P7→8.1 path end-to-end on real dual-layer FEL samples (both the FEL-safe and FEL-complex branches of the enhancement-layer classifier, plus encode-preserve), and adds a non-blocking advisory on copy paths (stream-copy / Remux / Trim & Concat): a P7 source copies 1:1, but TVs and hardware players don't decode Profile 7 from files — the message points to Transform RPU (P7→8.1) for universal playback - AV1 HW detection hardened + VFR awareness (v77) — the AV1-capable check now confirms real capability with a quick on-GPU probe across every backend (NVENC / AMF / VideoToolbox, joining QSV / VAAPI from v75) and on Windows, not just GPU-model matching — a future AV1-capable card the suite doesn't recognize by name is picked up automatically, no whitelist update needed (clean SW fallback if it can't encode AV1). Plus: on variable-frame-rate (VFR) sources, preserving dynamic HDR10+/DV on hardware now warns that metadata aligns positionally (tools may report a small tail frame-count difference, auto-adjusted, no visible impact) instead of staying silent — APV now keeps dynamic HDR10+ on small gaps too (tail-aligned like DV/HW), falling back to static HDR10 only on a large mismatch. Built-in DJI profiles + burn-in presets realigned to the current schema (test-guarded)
- More accurate HDR brightness measurement (v79) — the opt-in measure real MaxCLL/MaxFALL from video pass (used mainly for HLG→HDR10, where the source carries no light-level metadata) now takes the brightest of the three colors (max R/G/B) per pixel per CTA-861.3 instead of perceived luma, so highly saturated bright highlights (red/blue) are no longer underestimated and the written MaxCLL reflects the true peak. Still opt-in (
HDR10_MEASURE_CLL), one analysis pass, mastering display unchanged - Unified telemetry — DJI · GoPro GPMF · Sony NMEA · Garmin VIRB FIT · QuickTime ISO 6709 → one 24-column normalized CSV (v54) + SRT overlay tracks
- Richer telemetry (v54) — DJI full per-sample GPS track (fixes Osmo Action 6) with computed speed/heading + G-force; GoPro ACCL/GYRO + GPS9 (Hero 11/12/13); Garmin FIT enhanced speed/altitude + fixed temperature; GPS fix quality / sats / HDOP; modern KML
<gx:Track>import (Strava / Garmin / Google) - Metadata-only HDR/DV tools (v55/v56) — RPU profile transforms (force 8.1 / P5→8.1 / P7→8.1 dual-layer-aware / 8.1 preserving mapping / →P10 AV1) via the correct
editorpath, HDR10+→DV hybrid with source-derived L6 mastering metadata, aggregated RPU inspect summary +--verifyHDR10+ + RPU JSON export, remove DV / remove HDR10+ layers, plot DV L1/L2/L8 → PNG; HEVC + AV1, lossless on video. AV1 DV inject now works — auto-repairs the missing T.35 alignment byte thatav1dovi_tooldrops (dav1d-compatible; HDR10+ in hybrids untouched) and re-positions the metadata OBUs to the spec-compliant spot (v92: after all non-shown frames, immediately before the shown frame — matching real-world DV AV1 content; GPAC/MP4Box no longer flags the placement) - Mux tools (v49 + v50) — standalone script, 3 lossless flows: Remux (per-stream selection + per-target compat matrix), Demux (smart per-codec wrapping: video→
.mkv, audio→.mka, subs→native ext, chapters→Matroska XML), Mux (combine video + N audio/subs/chapters/attachments into a fresh container). Input: mkv/webm/mp4/m4v/mov/ts/m2ts/mts/vob/mxf - Spec-compliant HDR10/HLG output (v52 SW + v53 HW) — fixed a long-standing VUI signaling bug (streams reported
color_*=unknown, silently disabling x265hdr10-opt); now correct end-to-end across SW encoders and all 6 HW backends, via-x265-params/-svtav1-paramsplus post-encode bitstream filters - Rate control: CRF · 1-pass · 2-pass VBR (v51 + v53) — true 2-pass on SW encoders, NVENC
-multipass fullresonENCODE_MODE=3; automatic VBV/Level/Tier; HDR10 static metadata (Mastering Display + MaxCLL) injected on all PQ output, with opt-in measured MaxCLL/MaxFALL (v63) - Per-track audio control (v67–v68) — on multi-track files, choose per track: re-encode / copy lossless / drop, with bitrate + downmix computed per track's channel count; correct multi-codec re-encode order so a chosen codec actually applies (v66);
AV_AUDIO_TRACKS/AV_AUDIO_DROPenv for non-interactive selection; container-compat warnings when a copied track can't fit the target container (encode + concat); smart stream-copy keeps your per-track audio choice on every platform. Single-track files are untouched AV_PROFILEenv var (v52) — non-interactive profile auto-load for CI/cron/batch; resolvesUserProfiles/→profiles/*/, EXTENDS + schema validation preserved- Batch resume + recursive folders + profile system — quit anytime, re-run, picks up where it stopped;
.confprofiles withEXTENDSinheritance and schema validation - DJI Osmo Action 6 presets shipped — Airsoft Indoor/Outdoor · Moto Outdoor/Cinematic · D-Log M with LUT
- Burn-in overlay suite (v48 + v58 HDR-aware) — 4 flows: telemetry HUD (Python+matplotlib, 3 presets + map M2), SRT/ASS (libass styling), Image subs PGS/VobSub (external + embedded); detects source HDR/HDR10+/HLG/LOG/DV per file and proposes the right path (preserve / tonemap / brand LUT / refuse on DV);
BURNIN_HDR_POLICYenv for batch bypass; opt-in 5s preview. v80: configurable overlay font (FONT_FAMILY— family name or.ttf/.otfpath, applied to all overlay text) + three new corner readouts — altitude / heading / temperature (HUD_ALTITUDE/HEADING/TEMPERATURE, shown when the data-strip is off so they never duplicate it). v81: instant still-layout preview on the HUD flow — instead of a full render (or the 5s clip), composite one HUD frame over a real video frame from mid-clip, optionally with a positioning grid, to check element placement and readability fast (<name>_preview.png, auto-opened). v82: the still preview is now tonemapped on HDR-preserve sources so it no longer looks dim (only the preview PNG — the real encode keeps HDR;BURNIN_STILL_NO_TONEMAP=1for the raw frame) + optional text shaping for complex scripts (Arabic / Hebrew / Indic) on SRT/ASS; the broken ASS font-scale options (1.25x/1.5x, dead since v48) were removed since ASS files carry their own styling. v84: a visual HUD layout designer in your browser — pick a video (with its telemetry CSV, or demo mode with synthetic data), drag the HUD elements (data strip, map, corner readouts) directly over real video frames rendered by the same production engine (true WYSIWYG, HDR frames tonemapped for display), scrub the timeline, tune fonts/colors/sizes live, then save the layout as a custom preset (UserProfiles/burnin/) selectable in the HUD flow; runs strictly local (127.0.0.1), zero new dependencies
# Termux / Linux / macOS
chmod +x src/*.sh src/tools/*.sh
cd src && ./av_launcher.sh# Windows
Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned # once
cd src; .\av_encode.ps1Drop your files into the input folder shown on first run (Termux: /storage/emulated/0/Media/InputVideos/; Linux/macOS/Windows: InputVideos/ next to the script). Pick 1) Encode from the menu, accept the defaults — done. Output lands in the matching OutputVideos/ folder, with a per-batch summary on screen.
| Video | HEVC · H.264 · AV1 (SVT-AV1 / libaom) · DNxHR · ProRes · APV |
| HW backends | NVENC · QSV · VAAPI · VideoToolbox · AMF · MediaCodec |
| HDR | HDR10 · HDR10+ · Dolby Vision (P5/P7/P8.1 HEVC + P10 AV1) · HLG · LOG |
| Audio encode | AAC · Opus · FLAC · E-AC3 · AC3 · LPCM · Eclipsa/IAMF authoring incl. 7.1.4 + Atmos→Eclipsa render via Cavern (v88/v89) · per-track encode/copy/skip + container-compat warnings (v67–v68) |
| Audio passthrough | TrueHD · DTS · DTS-HD (auto stream copy; AC3 source copies too) · Atmos/DTS:X guarded (v87) · Eclipsa/IAMF group re-grafted (v88) |
| Telemetry | DJI · GoPro GPMF · Sony NMEA · Garmin VIRB FIT · QuickTime ISO 6709 |
| Workflows | Encode · Audio-only · Trim & Concat · Remux · HDR/DV tools · Telemetry · GPS · Burn-in (HUD/SRT/ASS/Image) · Analysis |
| Platform | Scripts | Setup |
|---|---|---|
| Termux (Android) | .sh |
pkg install ffmpeg termux-api exiftool python |
| Linux | .sh (bash 4+) |
distro pkg: ffmpeg python3 exiftool |
| macOS | .sh (bash 4+) |
brew install bash ffmpeg python3 exiftool |
| Windows | .ps1 (PS 5.1+) |
FFmpeg from gyan.dev in PATH |
detect_platform() discriminates Termux/Linux/macOS at startup; wrappers abstract GNU vs BSD coreutils, wake-lock and notifications.
| Encoder | Bitdepth | HDR10 | HDR10+ | DV | HLG | LOG |
|---|---|---|---|---|---|---|
| libx265 | 8/10 | ✅ | ✅ inline | ✅ 8.1 preserve/convert/copy (v45) | ✅ | ✅ |
| libx264 | 8/10 | ⚠ basic | — | — | ⚠ native only | ✅ |
| SVT-AV1 | 10 | ✅ | ✅ inline (v1.5+) | ✅ P10 preserve (v44) | ✅ | ✅ |
| libaom-av1 | 10 | ✅ | fallback HDR10 | — | ✅ | ✅ |
| DNxHR | 8/10 | ✅ preserve¹ | — | — | ✅ preserve¹ | ✅¹ |
| ProRes KS | 10/12 | ✅ preserve | — | — | ✅ preserve | ✅ |
| APV (FF 8.1+) | 10/12 · 4:2:2/4:4:4 (+alpha) | ✅ preserve | ✅ per-frame (v69) | — | ✅ preserve | ✅ |
¹ DNxHR HQX/444 = 10-bit (HDR); LB/SQ/HQ = 8-bit (SDR/proxy — lower precision on LOG/HDR, warned). Mezzanine codecs (DNxHR/ProRes/APV) preserve the HDR10/HLG picture + color tags but do not carry Dolby Vision RPU → honest per-source warnings (DV → clean HDR10 base; use x265/AV1 to keep DV). HDR10+ dynamic metadata is lost on DNxHR/ProRes — but preserved on APV since v69 (suite engine, see Highlights). ProRes 4444 XQ = native XQ profile (v64). ProRes (v74): analysis shows the exact profile (Proxy…4444 XQ), 4444/XQ add an alpha channel only when the source has one, container MOV or MXF; runs on all platforms via software prores_ks (VideoToolbox HW is macOS-only). DNxHR (v74): analysis shows the exact profile (LB/SQ/HQ/HQX/444). DNxHR MXF audio = PCM only (v64). APV (v65) = liboapv/libopenapv auto-detect, independent profile/preset/QP knobs (x265-style), mp4/mov/mkv; full bash + PowerShell parity. APV (v74): analysis shows the exact profile (4:2:2 / 4:4:4 / 4:4:4:4 · 10/12-bit) + new 4:4:4-alpha 12-bit (4444-12) profile.
DV preserve: extract source RPU codec-aware (dovi_tool HEVC / av1dovi_tool AV1) → re-encode HDR10 base → post-encode RPU inject → re-mux audio. Container DV signaling (DOVI configuration record) is written when the tool is available — HEVC: MKV via mkvmerge (v70) + MP4/MOV via MP4Box (v71); AV1: MKV via mkvmerge (v71) + MP4/MOV via MP4Box (v72) → DV recognized by TVs, not just PC players (silent in-stream-only fallback when absent). All common combinations (HEVC/AV1 × MKV/MP4/MOV) are covered.
| Backend | Platform | H.264 | H.265 | AV1 (HW) | HDR10 | Notes |
|---|---|---|---|---|---|---|
| NVENC | Win + Linux | ✅ | ✅ | RTX 40+ / Ada | ✅ 10-bit | p1..p7 |
| QSV | Win + Linux | ✅ | ✅ | Arc / Core Ultra / Meteor+ | Tiger Lake+ | veryfast..veryslow |
| VAAPI | Linux | ✅ | ✅ | Arc / Core Ultra / Meteor+ | ✅ | /dev/dri/renderD*, -quality 7..1 |
| VideoToolbox | macOS | ✅ | ✅ + ProRes | M3+ Apple Silicon | AS only (+ HLG) | q:v 50..80 |
| AMF | Win + Linux exp (v42.1) | ✅ | ✅ | RDNA3+ (RX 7000/8000 + 740M–890M) | ✅ | speed/balanced/quality |
| MediaCodec | Termux/Android | ✅ | ✅ + HDR10 repair (8-bit input) | none on mobile → SW fallback | via hevc_metadata bsf |
60%..150%, SoC whitelist |
Uniform preset table 1..7 (Ultrafast → Veryslow, default 4=Quality) across all backends. show_hdr_hw_dialog per-source-type (DV / HDR10+ / HLG / HDR10) with sw_full / sw_degraded / hw_hdr10 / hw_hlg / hw_sdr / hw_preserve (v46, DV preserve via HW) / skip. Bypass via HW_HDR_POLICY.
v75 — HW correctness audit (all 6 backends): VAAPI -quality + VideoToolbox q:v preset direction was inverted vs slot → fixed (NVENC/QSV/AMF were already correct). Intel AV1 (QSV/VAAPI) now gated by GPU model — older Intel that lists av1_qsv but fails at runtime (UHD) is no longer offered → clean SW fallback. NVENC quality path uses constant-quality VBR (-rc vbr -cq); AMF sends explicit usage=transcoding (+ main profile on AV1); PS1 HW warns on HDR10+ (dynamic metadata dropped → HDR10 static, opt-skip). MediaCodec: flagship Snapdragon (QTI vendor) + modern Exynos (s5e codename) now recognized as [verificat]; 10-bit HDR is signaled/containered but 8-bit input precision (ffmpeg MediaCodec encoders accept only 8-bit — validated on SD 8 Gen 3 + Exynos 1380). AV1 HW encode no longer claimed on any SoC — mobile AV1 hardware encode is near-nonexistent (decode is widespread; encode absent on SD 8 Gen 3 + Exynos 1380, both verified) → MC_CAP_AV1=0 everywhere, so av1_mediacodec falls back to the suite's libsvtav1 (better quality/control than the hidden MediaCodec SW libaom; HW_FORCE=1 to override). Encoder presence ≠ HW: MediaCodec falls back to SW transparently.
v77 — AV1 HW probe extended to all backends: the "real capability > model" principle (v75, Intel-only) now covers NVENC / AMF / VideoToolbox (bash) and every vendor on Windows (Get-GPUCapabilities). A short model || probe runs only when the GPU name isn't recognized (zero cost on known cards) — so a future AV1-capable GPU is detected at runtime without a whitelist update, and a listed-but-incapable encoder still falls back cleanly to SW. Probe can't false-positive (an incapable encoder fails the micro-encode clearly).
v46 — HW DV preserve: all 6 HW backends now support DV preserve (HEVC target via Profile 8.1, AV1 target via Profile 10). Extracts source RPU → HW encode HDR10 base → post-encode RPU inject. MediaCodec adds SEI signaling repair between encode and inject. Available on 12 combinations (6 backends × 2 codecs).
| Format | Encode | Stream copy | Cross-codec | Transform-only |
|---|---|---|---|---|
| HDR10 | all encoders | ✅ | HEVC↔AV1 | — |
| HDR10+ dynamic | HEVC + SVT-AV1 v1.5+ · APV (v69) | ✅ | all directions incl. APV↔HEVC/AV1 | synthesize DV from HDR10+ (v45) |
| Dolby Vision | HEVC 8.1 (v45) · AV1 P10 (v44) | ✅ | HEVC↔AV1 (v44/v45) | profile transform 5/7→8.1, 8.1↔10 · container signaling: HEVC + AV1 in MKV & MP4/MOV (mkvmerge / MP4Box, v70–v72) |
| HLG | all encoders + HW + MediaCodec (v39) | ✅ | ✅ | — |
| LOG (Apple/D-Log M/Samsung) | .cube LUT → Rec.709 / HLG · keep LOG |
✅ | — | — |
| SDR tonemap | Hable zscale → Rec.709 | — | — | — |
- Triple-layer hybrid — DV 8.1 (HEVC) or DV P10 (AV1) + HDR10 base + HDR10+ dynamic in one bitstream; v45 preserves both layers when source has DV + HDR10+ embedded
- HDR10+ → DV hybrid (v45, transform-only) — extract HDR10+ JSON → synthesize DV RPU → inject → re-mux; 100% lossless on video
- HDR10 → HLG (v63) — encode-time conversion (HEVC + AV1), symmetric with HLG → HDR10; HLG is metadata-free and plays gracefully on SDR screens (broadcast/phone-friendly)
| Codec | Default | Surround auto | Container |
|---|---|---|---|
| AAC | 192k | 5.1 → 384k · 7.1 → 768k | all |
| Opus | 128k | 5.1 → 256k · 7.1 → 512k | mkv/webm |
| E-AC3 | 224k stereo | 5.1 → 640k · 7.1 → 1024k | ❌ MOV |
| AC3 legacy (v53) | 224k stereo | 5.1 → 448k (max 640k) · 7.1 → 5.1 downmix | ❌ MOV |
| FLAC lossless | level 8 | passthrough | mkv only |
| LPCM | 16/24/32 le | passthrough | all |
| Eclipsa (IAMF) (v88/v89) | Opus 256k substreams | layout follows source (2/6/8/12 ch); Atmos sources can render objects → 7.1.4 via Cavern (v89) | mp4/mov only (needs MP4Box) |
| Codec | Behaviour |
|---|---|
| AC3 | preserved 1:1 |
| TrueHD | preserved 1:1 (+ HD warning) |
| DTS / DTS-HD MA | preserved 1:1 (+ HD warning) |
| Dolby Atmos / DTS:X (v87) | copy-first guard (objects survive only a 1:1 copy) + MP4/MOV Atmos container signaling |
| Eclipsa / IAMF (v88/v90) | group re-grafted onto MP4/MOV output after any 1:1 flow, incl. Burn-in full outputs (v90) — ffmpeg alone flattens it |
Loudnorm EBU R128 two-pass to -24 LUFS, optional per batch.
| Brand | Detection | Parser | Norm CSV | SRT |
|---|---|---|---|---|
| DJI | djmd/dbgi |
ExifTool (per-sample) | ✅ full track + computed speed/heading + G-force | ✅ |
| GoPro | gpmd |
GPMF KLV (Python) | ✅ | ✅ |
| Sony Action Cam | nmea |
NMEA 0183 (Python) | ✅ | ✅ |
| Garmin VIRB | fdsc |
FIT (Python) | ✅ + HR/cadence/power | ✅ |
| QuickTime | com.apple.quicktime.location.ISO6709 |
ExifTool | ✅ (single point) | ✅ |
- Norm CSV (24 cols, v54) —
timestamp,lat,lon,alt_m,speed_mps,speed_kmh,heading_deg,gforce_xyz,gyro_xyz,temp_c,hr_bpm,cadence_rpm,power_w,pitch_deg,roll_deg,yaw_deg,fix_quality,num_sats,hdop,source_brand - DJI strip sub-modes — strip telemetry · strip telemetry + cover · keep native GPS (drop cover only, re-graft djmd via MP4Box, v78)
- Excluded by design — Insta360, Yi/Akaso/SJCAM
- External GPS import — GPX/FIT/KML → CSV/SRT/GPX/KML via
av_extractor_gps - Embed lossless (opt 7) — extract telemetry then re-mux as SRT subtitle track + attachments (MKV); video stream copy, no re-encode. Submenu with 4 profiles:
srt(SRT only, any container) ·srt_csv(SRT + norm CSV, MKV) ·srt_csv_gpx(default, SRT + norm CSV + GPX, MKV) ·all(SRT + basic + FULL + GPX + KML, MKV mandatory). KML auto-generated for profileallfrom norm CSV (or DJI exiftool template). Mimetypes:text/csv,application/gpx+xml,application/vnd.google-earth.kml+xml.
-
Encode — per-file HDR/DV/HLG/LOG dialogs; recursive scan with
PRESERVE_FOLDER_STRUCTURE=1 -
Audio-only re-encode — video stream copy + audio re-encode
-
Media analysis (
av_check) — 30-50 field CSV -
Telemetry — 6 modes: Standard / Full / SRT / All / Raw / Strip
-
External GPS — GPX/FIT/KML → CSV/SRT
-
Trim & Concat — single trim · concat (auto demuxer/filter) · pipeline 3-pass (+ dry-run preview, v63) · batch trim · HDR/LOG-aware re-encode (v60: per-file dialog — preserve HDR10/HLG · tonemap · LUT · keep LOG · skip; pipeline x265 + svtav1 HDR, codec_tag); v73: LOG also detected at concat (brand LUT / keep LOG / skip) + DV container signaling kept on stream-copy / pipeline smart-copy / audio-only paths; v86: honest message when re-encode is forced by variable frame rate (VFR) sources — lossless concat on VFR cuts would corrupt timestamps at the joins, so the compat check correctly re-encodes and now says exactly why (CFR sources still concat lossless)
-
Mux tools (v49 + v50, no re-encode, lossless):
Flow Action Remux container Per-stream selection (video/audio/sub/attach/chapters) → mkv/mp4/mov/webm with per-target compat matrix + preflight Demux streams Smart per-codec wrap: video→ .mkv, audio→.mka, sub→native ext, cover→jpg/png, chapters→XML, attachments+data→foldersMux streams (v50) Combine video + N audio + N subs + chapters + attachments → fresh container. Manual selection, VobSub pair handling, per-track metadata opt-in (lang/title/default/forced). Compat matrix reused from Remux. Input only from InputVideos/. -
HDR/DV Tools (v44/v45/v56, metadata-only, no re-encode):
Option Action Transform RPU DV profile convert (5→8.1, force 8.1, 8.1 preserving mapping, →P10 AV1) Inspect ffprobe + dovi_tool info -s+ RPU JSON export + HDR10+--verifyHDR10+ → DV hybrid (v45) Synthesize DV RPU from HDR10+ → inject → re-mux Remove DV → HDR10 (v56) Strip DV layer, keep HDR10/HDR10+ (HEVC + AV1) Remove HDR10+ (v56) Strip HDR10+ metadata, keep HDR10/DV (HEVC + AV1) Plot DV metadata (v56) L1/L2/L8 brightness/trims → PNG (native) -
Burn-in overlay (v48, re-encode with baked-in pixels):
Flow Source ffmpeg pipeline Telemetry HUD <name>_norm.csvburnin_render.py→ PNG seq →[0:v][1:v]overlaySRT <name>.srtsubtitles='<path>':force_style='FontSize=N,...'[:shaping=<mode>](libass)ASS <name>.assass='<path>'[:shaping=<mode>](libass — ASS carries its own styling)Image subs <name>.sup·<name>.idx+.sub· embedded PGS/VobSub[0:v][1:s]overlay(ext) ·[0:v][0:s:N]overlay(emb)Text shaping (v82, SRT/ASS): optional
auto/simple/complexfor complex scripts (Arabic/Hebrew/Indic), gated on ffmpeg support;auto= unchanged default.Preview mode (opt-in per flow): HUD offers a still-layout preview (one composited frame at mid-point, tonemapped on HDR-preserve sources so it isn't dim — v81/v82) or a 5s clip; SRT/ASS/Image offer the 5s clip. Output
<name>_preview.<ext>(.pngfor the still) — a quick check before committing to full encode.
AV-Encoder-Suite/
├── src/
│ ├── av_launcher.sh # Interactive main menu (Termux/Linux/macOS)
│ ├── av_common.sh # Shared functions (HDR, DJI, audio, progress, HW dispatch)
│ ├── av_encoder_x265.sh # H.265/HEVC encoder
│ ├── av_encoder_x264.sh # H.264/AVC encoder
│ ├── av_encoder_av1.sh # AV1 encoder (SVT-AV1 / libaom)
│ ├── av_encoder_dnxhr.sh # DNxHR encoder (Avid mezzanine)
│ ├── av_encoder_prores.sh # ProRes encoder (Apple professional)
│ ├── av_encoder_apv.sh # APV encoder (Samsung, ffmpeg 8.1+)
│ ├── av_encoder_audio.sh # Audio-only re-encode (video stream copy)
│ ├── av_trimconcat.sh # Trim & Concat pipeline (v36/v37 + v60 HDR/LOG, standalone)
│ ├── av_mux.sh # v49+v50 — Mux tools standalone (remux + demux + mux)
│ ├── av_mux.ps1 # v49+v50 — PS1 mirror standalone
│ ├── av_hdr_dv_tools.sh # HDR/DV tools submenu (v44/v45/v56; remux moved to av_mux in v49)
│ ├── av_check.sh # Media analysis + CSV export (Termux/Linux/macOS)
│ ├── av_check.ps1 # Media analysis + CSV export (Windows)
│ ├── av_encode.ps1 # All-in-one PowerShell script (Windows)
│ ├── av_telemetry.sh # Unified telemetry extractor (5 brands)
│ ├── av_telemetry.ps1 # PS1 mirror standalone
│ ├── av_extractor_gps.sh # External GPS import GPX/FIT/KML
│ ├── av_extractor_gps.ps1 # PS1 mirror standalone
│ ├── av_burnin.sh # v48 — Burn-in overlay (HUD/SRT/ASS/Image)
│ ├── av_burnin.ps1 # v48 — PS1 mirror standalone
│ ├── burnin_render.py # v48 — Python+matplotlib HUD render engine
│ ├── burnin_designer.py # v84 — local web server for the visual HUD layout designer
│ ├── burnin_designer.html # v84 — designer UI (single page, vanilla JS, served locally)
│ ├── av1_dv_t35_repair.py # v56 — AV1 DV T.35 trailing-byte repair (dav1d compat) + v92 spec-compliant OBU reorder
│ ├── dji_djmd_dlogm.py # v62 — DJI Action 6 D-Log M detector (djmd protobuf .2.4.1==19)
│ ├── apv_hdr10plus.py # v69 — APV HDR10+ engine (ST 2094-40 inject/extract/probe)
│ ├── dv_p7_analyze.py # v76 — DV Profile 7 enhancement-layer classifier (MEL / FEL-safe / FEL-complex)
│ ├── burnin_presets/ # v48 — HUD layout presets (.conf)
│ │ ├── minimal.conf # timestamp + speed corner overlay
│ │ ├── data-strip.conf # bottom bar gauges (speed/alt/heading/temp)
│ │ └── full.conf # data-strip + map M2 + G-force/HR gauges
│ ├── profiles/
│ │ ├── example_profile.conf # Documented profile example (all fields)
│ │ └── dji_action6/ # DJI Osmo Action 6 preset profiles
│ │ ├── DJI_Action6_Airsoft_Indoor.conf
│ │ ├── DJI_Action6_Airsoft_Outdoor.conf
│ │ ├── DJI_Action6_Moto_Outdoor.conf
│ │ ├── DJI_Action6_Moto_Cinematic.conf # D-Log M + LUT
│ │ └── DJI_Action6_DLogM_Outdoor.conf # D-Log M + LUT
│ └── tools/
│ ├── hdr10plus_parser.sh # hdr10plus_tool HEVC installer (quietvoid; Rust)
│ ├── hdr10plus_parser.ps1 # hdr10plus_tool HEVC installer (Windows, prebuilt)
│ ├── av1hdr10plus_parser.sh # v44 — sven-pke fork installer (AV1 OBU); binary `av1hdr10plus_tool`
│ ├── av1hdr10plus_parser.ps1 # v44 — sven-pke fork installer (Windows, cargo build)
│ ├── dovi_parser.sh # dovi_tool HEVC installer (quietvoid; Rust)
│ ├── dovi_parser.ps1 # dovi_tool HEVC installer (Windows, prebuilt)
│ ├── av1dovi_parser.sh # v44 — sven-pke fork installer (AV1 RPU); binary `av1dovi_tool`
│ ├── av1dovi_parser.ps1 # v44 — sven-pke fork installer (Windows, cargo build)
│ ├── exiftool_update.sh # ExifTool smart updater (Termux/Linux/macOS)
│ ├── exiftool_update.ps1 # ExifTool smart updater (Windows)
│ ├── openapv_validator.sh # v69 — OpenAPV reference decoder installer (cmake build)
│ ├── openapv_validator.ps1 # v69 — OpenAPV reference decoder installer (Windows, prebuilt)
│ ├── mkvmerge_installer.sh # v70 — mkvmerge installer (pkg manager; dvcC DV on MKV, HEVC+AV1)
│ ├── mkvmerge_installer.ps1 # v70 — mkvmerge installer (Windows, portable .7z)
│ ├── mp4box_installer.sh # v71 — MP4Box/GPAC installer (pkg manager; dvcC DV on MP4/MOV)
│ ├── mp4box_installer.ps1 # v71 — MP4Box/GPAC installer (Windows, winget)
│ ├── cavernize_installer.sh # v89 — Cavernize guide (macOS manual; Linux/Termux unavailable)
│ ├── cavernize_installer.ps1 # v89 — Cavernize installer (Windows, official portable zip)
│ ├── profile_diff.sh # v43 — compare two .conf profiles (bash)
│ └── profile_diff.ps1 # v43 — compare two .conf profiles (PS1 mirror)
├── docs/
│ ├── av_info.txt # Full setup & usage documentation
│ └── av_changelog.txt # Version history
├── tests/ # bash + PS1 test framework (v43)
│ ├── framework.sh / .ps1 # Assertion library
│ ├── run_tests.sh / .ps1 # Discover + run test_*.{sh,ps1}; per-test log + summary
│ ├── _helpers.ps1 # AST-based function importer for PS1 tests
│ ├── fixtures/
│ │ ├── generate_samples.sh / .ps1 # ffmpeg-driven synthetic samples (idempotent)
│ │ └── samples/ # Generated samples (gitignored)
│ ├── unit/ # Pure-logic tests (no ffmpeg required)
│ ├── integration/ # ffmpeg/ffprobe/python-dependent (auto-skip on missing deps)
│ └── results/ # Per-test logs (gitignored)
├── .gitignore
├── LICENSE
└── README.md
bash tests/fixtures/generate_samples.sh && bash tests/run_tests.sh
.\tests\fixtures\generate_samples.ps1; .\tests\run_tests.ps1Tests dependent on ffmpeg/ffprobe/python3/exiftool auto-skip when the binary is missing.
- Encode video + audio
- Encode audio only (video stream copy)
- Analyze media (CSV export)
- Telemetry — DJI / GoPro / Sony / Garmin / QuickTime
- Import external GPS — GPX/FIT/KML
- Trim & Concat — trim / concat / pipeline / batch
- Mux tools (v49 + v50) — Remux + Demux + Mux, no re-encode · 10 input formats (Blu-ray/DVD/broadcast) · remux to mkv/mp4/mov/webm · demux to mkv/mka/native sub ext + cover/chapters/attach · mux (v50): combine separate video + audio[N] + subs[N] + chapters + attachments into fresh container (manual selection, VobSub pair handling, per-track metadata opt-in)
- HDR/DV tools — transform / inspect / HDR10+→DV / remove DV / remove HDR10+ / plot (v56)
- Burn-in overlay (v48) — HUD telemetry / SRT / ASS / Image subs (PGS/VobSub)
- Exit
KEY=VALUE files (bash source / PS1 Get-Content); user profiles in UserProfiles/. Built-in DJI Osmo Action 6 presets (airsoft indoor/outdoor, moto outdoor/cinematic, D-Log M). Schema validation per key (enum / regex / int / path) before load; unknown keys are forward-compat warnings.
EXTENDS inheritance — EXTENDS=parent_name resolves through sibling dir → UserProfiles/ → src/profiles/*/; root→leaf override; cycle detection + depth limit 5.
Bypass fields for batches: HW_HDR_POLICY, MEDIACODEC_HDR_POLICY, DOVI_PRESERVE_POLICY (v45), HW_FORCE.
profile_diff CLI — src/tools/profile_diff.{sh,ps1} A.conf B.conf — exit 0 identical / 1 different / 2 invalid.
| Platform | Video | CSV | Telemetry |
|---|---|---|---|
| Termux | /storage/emulated/0/Media/OutputVideos/ |
av_check_report.csv |
.gpx / _basic.csv / _FULL.csv / _norm.csv / .srt |
| Linux / macOS | $SCRIPT_DIR/OutputVideos/ |
same | same |
| Windows | output\ next to script |
same | same |
| Tool | Purpose | bash | PS1 |
|---|---|---|---|
hdr10plus_tool |
HDR10+ HEVC (quietvoid) | hdr10plus_parser.sh |
hdr10plus_parser.ps1 (prebuilt) |
av1hdr10plus_tool |
HDR10+ AV1 (sven-pke fork) | av1hdr10plus_parser.sh (cargo) |
av1hdr10plus_parser.ps1 (rustup + git) |
dovi_tool |
DV RPU HEVC (quietvoid) | dovi_parser.sh |
dovi_parser.ps1 (prebuilt) |
av1dovi_tool |
DV RPU AV1 (sven-pke fork) | av1dovi_parser.sh (cargo) |
av1dovi_parser.ps1 (rustup + git) |
| ExifTool updater | Latest version | exiftool_update.sh |
exiftool_update.ps1 |
oapv_app_dec |
OpenAPV reference decoder — optional APV decode-check (v69) | openapv_validator.sh (cmake) |
openapv_validator.ps1 (prebuilt) |
mkvmerge |
MKVToolNix — optional, DV container signaling (dvcC) on MKV for HEVC (v70) + AV1 (v71) | mkvmerge_installer.sh (pkg manager) |
mkvmerge_installer.ps1 (portable .7z) |
MP4Box (GPAC) |
optional, DV container signaling (dvcC) on MP4/MOV for HEVC (v71) + AV1 (v72) | mp4box_installer.sh (pkg manager) |
mp4box_installer.ps1 (winget) |
Cavernize (Cavern) |
optional, renders Atmos objects → 7.1.4 channels for Eclipsa/IAMF authoring (v89); needs .NET Desktop 8+ | cavernize_installer.sh (macOS guide; Linux/Termux unavailable) |
cavernize_installer.ps1 (official portable zip) |
| Profile diff | Compare .conf files |
profile_diff.sh A B |
profile_diff.ps1 A B |
AV1 forks install with renamed binaries (av1dovi_tool, av1hdr10plus_tool) so they coexist with HEVC upstream. Since v69, every external tool name is overridable via AV_TOOL_* environment variables (plain name or full path) — e.g. AV_TOOL_DOVI=dovi-tool for distros that ship renamed binaries. Since v93, DLL-based portable packages are also auto-discovered when co-located next to the scripts — GPAC in src/GPAC/, Cavernize in src/cavernize/ (or src/tools/cavernize/, where the installer unpacks); resolution order: env override → co-located folder → PATH. Plain single-exe tools placed next to the scripts were already found automatically.
| Problem | Solution |
|---|---|
Permission denied on .sh |
chmod +x src/*.sh src/tools/*.sh |
command not found: ffmpeg |
pkg install ffmpeg / brew install ffmpeg / apt install ffmpeg / PATH (Win) |
| PS1 script blocked | Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned |
| Termux wake-lock fails | Install Termux:API from F-Droid (not Play Store) |
| DJI files lose GPS | Use MKV — DJI tracks incompatible with MP4/MOV |
| PGS/DVDSUB subs lost | Normal in MP4/MOV; use MKV |
| GPS export empty | DJI Action 6 has no internal GPS — needs DJI RC or DJI Mimo |
| macOS bash startup fails | Default 3.2 unsupported — brew install bash |
| AV1 HDR10+ inline missing | Requires SVT-AV1 v1.5+; libaom falls back to HDR10 static |
av1dovi_tool not found |
./tools/av1dovi_parser.sh (cargo build from sven-pke fork) |
MIT License — free to use, modify and distribute.
If you find this project useful, consider a small donation — it helps keep development going.
See docs/av_changelog.txt for full version history.
Current: v97 — 198 bugs fixed · 283+ features · ~63 950 LoC · 263 files
v97 highlights: the Eclipsa/IAMF authoring flow had been writing non-conformant files since v88. Each file was missing its mix description — the part that tells a player how to combine the channels. Nothing looked wrong: the file was created without error and the suite read its channel layout back correctly. The cause was a single character in the command sent to ffmpeg — a comma where a colon belonged — which made ffmpeg discard the entire mix block. ffmpeg did warn about it, but the message was hidden by the setting that only lets serious errors through. Players and ffmpeg up to 9.0 accepted such files, so the defect stayed invisible; ffmpeg 9.0 rejects them outright, as the IAMF specification requires at least one submix. Fixed across all four channel layouts (stereo, 5.1, 7.1, 7.1.4) on both platforms; the stereo case needed a second correction, a missing mix label on the element — which turns out to be the very reason the wrong form was chosen back in v88. The flow now also verifies the file it just produced: if the mix description is missing, the file is not delivered and the reason is stated, instead of trusting an exit code that in this case said nothing. Files authored with v88–v96 still play on current players but should be re-created before use with newer software. The same release also opens media analysis to audio files: option 3 used to skip anything without a picture, which meant the audio-only output of option 2 — Eclipsa included — could not be checked with the suite's own analyzer. Audio files are now reported in full (codec with its Atmos / DTS:X / Eclipsa label, bitrate, sample rate, depth, channels and layout, language, track count, subtitles, chapters, attachments), standalone audio formats are recognized (m4a, mka, flac, opus, wav, aac, ac3, eac3, dts, thd, iamf), and picture fields read N/A instead of inventing plausible-but-false values — previously an audio file would have shown an empty resolution, the whole-file bitrate reported as video bitrate, and "SDR" on something with no picture. The CSV keeps its 38 columns, so anything consuming the report still works. Suites: Linux 102/102 with no skips, Windows bash 100/0/2 of 102, Windows PowerShell 94/94.
v96 highlights: the bash side had always been tested on Windows through a compatibility layer; running it on native Linux for the first time surfaced things that environment could not show — there, filename case matters and a Windows executable simply won't start. Two bugs had been hiding since v71: MP4Box was never found on Linux/macOS (it is installed under that exact name, with capitals, while the suite looked for the lowercase form), which silently removed Dolby Vision signaling in MP4/MOV, Eclipsa authoring, container-level Atmos marking and native DJI GPS preservation — every flow degrades gracefully, so nothing ever looked broken; and on a working tree shared with Windows (WSL, network share, dual-boot) the suite picked the Windows executable sitting next to the scripts instead of the native tool, because the file exists and looks executable. The Dolby Vision / HDR10+ installers also never installed anything on Linux or macOS — they printed advice that cannot work and exited as if successful, while the code that actually installs was already there, reserved for Android; they now build and install on all three systems, fall back to a bundled font when the system font library is missing, report placement failures instead of claiming success, and point at a source build when a distribution has dropped the GPAC package (Ubuntu 26.04). The same run also closed a long-standing gap on the other side of the same coin: decimal numbers were formatted according to the machine's regional settings, so on a Linux with a European locale the burn-in preview and Trim & Concat thumbnails failed outright (ffmpeg rejects 1,504 as a timestamp), the CSV report came out with commas in five columns, and frame-rate conversion was silently skipped. Windows had been protected against this class since earlier versions; the Linux/macOS side had been assumed safe and never tested. Suites: Linux 101/101 with no skips, Windows bash 99/0/2 of 101, Windows PowerShell 93/93.
v95 highlights: a cleanup release — no behaviour changes at all. Nine functions with no callers were removed (375 lines): five from the shared bash code, four from the Windows script, where a whole Remux block had been left behind when that flow moved into its own script back in v49. Two of them were the bash/Windows pair of the same capability, dead on both platforms — the capability itself is alive and tested, just living elsewhere now. Tests that exercised the dead code were re-pointed at the copy that actually runs where one existed, and removed where the function was gone for good. A permanent guard now flags any function without callers, on both platforms; it self-checks against planted cases, ignores names that appear only in comments (which is how one of the nine hid from the first sweep), and judges per file on Windows, since scripts there are standalone by design — the exact reason two copies of the same helper had been allowed to drift apart. Two unused functions were deliberately kept: they belong to the cross-platform compatibility layer that another rule requires everyone to go through.
v94 highlights: a full audit campaign — no new features; every flow was driven through the real menus, on real files, on both platforms. Sixteen bugs fixed, six of which meant the flow produced nothing at all: x265 refused to start on ordinary resolutions (1080p60 / 1440p / 2.7K / any vertical footage) because the level was derived from image width instead of samples + sample rate + max dimension — now table-driven per spec, and informational-only on CRF · any path containing a space broke encoding on Windows (process launch joined arguments without quoting, so my clip.mp4 or a My Videos folder split the ffmpeg command) · AV1 VBR never worked on SVT-AV1, 1-pass or 2-pass, because the suite always sent a bitrate ceiling that SVT rejects outside CRF mode · AV1 "2-pass" was silently two 1-pass runs on builds where the stats syntax is unimplemented (info-level message, exit 0, no stats file) — now probed by result, not by version string · Opus failed outright on 5.1(side) sources (typical for AC3/DTS) · Concat MP4→MKV, the default combination, failed on the source timecode track. Plus honest reporting where it was wrong: the hybrid DV+HDR10+ completion label no longer claims an HDR10+ layer that isn't there, ProRes/APV keep the source colour signalling, single-frame preview actually lands mid-clip, and piped/scripted runs now get the announced menu defaults instead of empty values. Validation covered the metadata matrix end-to-end through the menus (HDR10+ and Dolby Vision in every direction incl. cross-codec, three-layer hybrid in MKV and MP4, Profile 7 → 8.1 on encode, HDR10+ on APV in all six combinations — where bash and PowerShell produced byte-identical bitstreams, metadata included).
v93 highlights: the suite is now fully self-contained on Windows — DLL-based portable tool packages are auto-discovered from folders sitting next to the scripts (GPAC/MP4Box from src/GPAC/, Cavernize from src/cavernize/ or src/tools/cavernize/), with AV_TOOL_* env still the supreme override and PATH the final fallback; no absolute paths anywhere (dev-box paths removed from the test suite too, which re-enabled a silently-gated real Cavern render test) · toolchain revalidated against the 2026-07-20 FFmpeg git build — all 398 upstream commits reviewed, zero suite changes needed; premises re-confirmed empirically (movenc still writes no Atmos/JOC dec3 extension so the v87 signaling stays required, mainline SVT-AV1 4.2.0 still lacks hdr10plus-json, IAMF still flattens on any mux, raw HEVC→MKV still refused) and the new upstream side-data types (SMPTE 2094-50, itut_t35 tracks) cause no false positives in detection; both full suites green on the new build.