Measured FFmpeg-v2 workflow improvements
The released M4 Max native profile reduced elapsed workflow time 22.61% for HD and 28.48% for live action/VFX against the prior Stock workflow preset. The workload includes planning, durable queueing, encode, metadata validation, complete strict decode, SHA-256 and staged no-clobber publication.
Before and after
| 60-second source | Stock prior preset, s | V2 native profile, s | Less elapsed time | Trials / configuration | Full quality frames |
|---|---|---|---|---|---|
| 1080p animation · 1920×1080 / 30fps | 9.7344 | 7.5339 | 22.61% | 9 | 1,800 |
| Live action / VFX · 1920×800 / 24fps | 7.5278 | 5.3840 | 28.48% | 9 | 1,440 |
Baseline: Homebrew Stock FFmpeg 9.0.2 with encoder/decoder/filter/validation pools 4/2/2/4. Candidate: native n9.0.2-3-gb6e0de6-ffmpeg-v2-m4 with 8/2/2/8. This changes tuning and uses an optional native source build; it is not a comparison against an automatically tuned or same-settings Stock default.
The earlier FFmpeg-v2 workflow is another reference: elapsed time decreased 22.31% on HD and 30.77% on live action. These reference labels must not be interchanged.
Quality stayed within every declared gate
Same source/range, libx264 veryfast/CRF20 and AAC160k. Quality was not traded for a lower configured quality/bitrate target. All target frames were aligned and checked. Output size increased slightly; this is not a compression-ratio improvement.
| Source | VMAF before → after | ΔPSNR-Y dB | ΔSSIM | Output size change |
|---|---|---|---|---|
| 1080p animation | 94.761184 → 94.781527 | +0.004890 | +0.000001 | +0.162% |
| Live action / VFX | 93.375096 → 93.379144 | +0.000316 | +0.000002 | +0.024% |
Gates: ΔVMAF≥−0.10, ΔPSNR-Y≥−0.05dB, ΔSSIM≥−0.00001, size≤1.02×, faster median and upper paired confidence ratio<1 against the declared baseline. Native/Stock outputs at matching pools were byte-identical on these source ranges.
Same-tuning Stock controls
Stock accepts the same thread controls. Applying the same 8/2/2/8 pools changes the question from packaged workflow improvement to engine comparison. That comparison does not establish a general native-encoder speed advantage.
| Source | V2 native, s | Matched Stock, s | V2 elapsed change | Ratio 95% interval |
|---|---|---|---|---|
| 1080p animation | 7.5339 | 6.8781 | +9.53% | 0.979218–1.315438 |
| Live action / VFX | 5.3840 | 5.4297 | -0.84% | 0.625907–1.070255 |
Positive change means more elapsed time. Both matched-control intervals include equality. The HD median is 9.53% slower; live action is 0.84% faster, without a proven engine advantage. The product benefit is the coordinated workflow policy plus job control, verification and integration.
Conditions and uncertainty
Apple M4 Max, 10P+4E, 36 GiB, macOS 27.2/arm64. Big Buck Bunny 1920×1080/30fps and Tears of Steel 1920×800/24fps, source range 60–120s. One excluded warmup then nine measured trials per configuration. Time reduction is one minus the ratio of total-time medians; 20,000 paired trial-block bootstrap draws supply uncertainty. Interpreter startup is excluded.
| Source | V2 / prior Stock ratio | Paired 95% interval |
|---|---|---|
| 1080p animation | 0.773946 | 0.577062–0.879356 |
| Live action / VFX | 0.715217 | 0.670815–0.765480 |
The host was shared, work was sequential and nice≥10, with a task-scoped system-memory waiver; own-footprint, disk, mount, timeout and output guards remained. Busy-host scheduling and small sample scope limit extrapolation. Times are historical released-profile measurements, not a new public 0.6.1 packaging benchmark. No cross-platform or universal FFmpeg speedup is claimed.
The auto profile applies only to eligible >=60s M4 Max native-geometry, progressive square-pixel, unrotated 8-bit software quality transcodes with explicit matching fps≤30. Other workloads retain conservative policies. See native builds and configuration.
Inspect the source evidence
Machine-readable released workflow summary · retained verification index · original detailed report.
Experimental PGO measurements are separate, remain unpromoted and are not added to these percentages. See the research preview in Results and the technical archive for failed and rejected outcomes.