FFMPEG-V2 / benchmarks

What improved over the Stock workflow?

22.61% less HD workflow time. 28.48% less live-action workflow time. The recorded released M4 Max native profile beat the prior Stock workflow preset while passing the declared quality and output-size gates.

22.61%

Less elapsed time · 1080p animation · prior Stock preset

28.48%

Less elapsed time · live action / vfx · prior Stock preset

Before and after

Complete workflow medians · 60s source range · nine trials per configuration
Source	Stock prior preset	V2 native profile	Less elapsed time	Quality
1080p animation	9.7344 s	7.5339 s	22.61%	Passed · 1,800 frames
Live action / VFX	7.5278 s	5.3840 s	28.48%	Passed · 1,440 frames

Exactly what is compared: Stock FFmpeg 9.0.2, encoder/decoder/filter/validation 4/2/2/4, versus the native M4 Max profile at 8/2/2/8. Both run the same planning, queue and full-validation contract. The native engine is an optional source build. This is a comparison with a prior workflow preset, rather than an automatically tuned Stock default.

Where the gain comes from

Coordinate the encoder and validation pools for eligible long software jobs. Retain metadata, full decode, hashing and staged publication. The earlier FFmpeg-v2 workflow is a separate baseline: time decreased 22.31% on HD and 30.77% on live action.

Quality and compression

Aligned full-frame checks · no reduced configured quality/bitrate target
Source	VMAF before → after	ΔPSNR-Y	ΔSSIM	Size change
1080p animation	94.761184 → 94.781527	+0.004890 dB	+0.000001	+0.162%
Live action / VFX	93.375096 → 93.379144	+0.000316 dB	+0.000002	+0.024%

All 1,800 HD and 1,440 live-action frames were checked against aligned source frames. The gates passed: ΔVMAF≥−0.10, ΔPSNR-Y≥−0.05dB, ΔSSIM≥−0.00001 and size≤1.02×. Output sizes increased slightly, so this is a time improvement rather than a compression-ratio claim. Same-pool Native/Stock outputs were byte-identical on the tested ranges.

Same tuning: the baseline matters

Stock FFmpeg accepts the same thread controls. With matched 8/2/2/8 pools, the engine comparison does not establish a general native-encoder speed advantage. FFmpeg-v2 adds the reusable tuning policy, job control, verification and integration around the engine.

Matched Stock control · positive change means more V2 elapsed time
Source	V2 native	Stock, same pools	V2 elapsed change	Ratio 95% interval
1080p animation	7.5339 s	6.8781 s	+9.53%	0.9792–1.3154
Live action / VFX	5.3840 s	5.4297 s	-0.84%	0.6259–1.0703

The HD median is 9.53% slower and the live-action median 0.84% faster; both confidence intervals include equality. This is why the headline names its prior-preset baseline.

Measured conditions

Apple M4 Max, 10P+4E, 36 GiB, macOS 27.2/arm64.
Big Buck Bunny 1920×1080/30fps and Tears of Steel 1920×800/24fps, 60–120s source ranges.
libx264 veryfast/CRF20, AAC160k; one excluded warmup and nine measured trials per configuration.
Ratio of workflow medians, with 20,000 paired trial-block bootstrap draws. Interpreter startup is excluded.
Eligible ≥60s native-geometry software quality jobs. Shared host and limited source/hardware scope; task-scoped memory waiver disclosed in the full method.

These are recorded released-profile measurements, not a rerun of public 0.6.1 packaging. Read the complete released-profile results and method, summary JSON, or the product comparison.

Experimental optimization results

Separate research candidate: x264 progress/PGO optimization, with matched quality on the target outputs. The candidate remains experimental and was not promoted. These percentages are not added to the released-profile gains above.

PGO candidate not promoted

Experimental complete HD workflow

Paired elapsed time decreased 1.743% / 1.791% against Stock, with nine valid blocks in each added-load state. This passes the bounded workflow groups.

Experimental complete HD workflow · accepted cases only
Workload / load	Reference	Valid / 9	Candidate s	Reference s	Paired change	95% ratio	Sample gate
hd60/load0	Original native	9 / 9	5.6018	5.6954	-1.764%	0.9791–0.9855	Pass, bounded
hd60/load0	Stock FFmpeg	9 / 9	5.6018	5.7016	-1.743%	0.9799–0.9853	Pass, bounded
hd60/load2	Original native	9 / 9	5.6307	5.7251	-1.532%	0.9822–0.9872	Pass, bounded
hd60/load2	Stock FFmpeg	9 / 9	5.6307	5.7376	-1.791%	0.9797–0.9849	Pass, bounded

Experimental direct HD and raw input

HD accepted 1/9 blocks in each state, raw 3/9 and 7/9. These groups have insufficient samples for release acceptance; a small-sample interval below 1 does not override that requirement.

Experimental direct encoder · accepted cases only
Workload / load	Reference	Valid / 9	Candidate s	Reference s	Paired change	95% ratio	Sample gate
hd60/load0	Original native	1 / 9	—	—	Insufficient	Not estimated	Insufficient
hd60/load0	Stock FFmpeg	1 / 9	—	—	Insufficient	Not estimated	Insufficient
hd60/load2	Original native	1 / 9	—	—	Insufficient	Not estimated	Insufficient
hd60/load2	Stock FFmpeg	1 / 9	—	—	Insufficient	Not estimated	Insufficient
raw30/load0	Original native	3 / 9	2.3782	2.4107	-1.654%	0.9803–0.9861	Insufficient
raw30/load0	Stock FFmpeg	3 / 9	2.3782	2.4121	-1.801%	0.9780–0.9844	Insufficient
raw30/load2	Original native	7 / 9	2.3285	2.3743	-2.033%	0.9777–0.9819	Insufficient
raw30/load2	Stock FFmpeg	7 / 9	2.3285	2.3755	-1.994%	0.9773–0.9829	Insufficient

These intervals use a geometric mean of paired whole-block ratios, distinct from the released-profile ratio-of-medians experiment. Warm-cache/load controls and the exact-quality gate apply; CPU instruction reduction is not wall-clock proof.

Research method and full quality table · Original detailed report · Research summary JSON · Direct block analysis · Workflow block analysis

Start using FFmpeg-v2 for a typed, supervised local workflow. The technical archive preserves all original methods, slowdowns, rejected experiments and remaining gaps.