v0.3.3

Follow-up release to chase down the last real-world VMAF failure mode:
animated content (The Croods et al.) was still scoring under 50 on
encodes that visually inspected as flawless. Adds proper stream
normalization in the VMAF filter graph, a whole-file compare for TV-
sized content, and a diagnostic cross-check so future suspicious scores
can be verified against independent metrics without guesswork. Also a
small UI wording fix for the VMAF-rejected case.

Changed

• VMAF filter graph now normalises frame rate and colour range on both streams. Source fps is probed up front and used as the target for an fps=fps=N clause on both the reference and distorted streams, which guarantees matching frame counts regardless of VFR/CFR mix. Both streams also run through scale=in_range=auto:out_range=tv so a source-vs-encode range-tag mismatch can't silently stretch luma values (the classic "looks fine, scores 49" signature on some WEBDL captures). The side-by-side probe summary (VMAF inputs — ref: 1920x1080 23.976fps yuv420p range=tv | dist: ...) is logged before every VMAF run so any future mismatch is visible at a glance.
• VMAF subprocess timeout now scales with analysed window (max(5 min, 3× duration)). A full-file compare on a 45-minute episode can't time out any more.

Added

• SSIM + PSNR cross-check on suspicious VMAF scores. Whenever VMAF reports under 80 on the mean, a second ffmpeg pass computes SSIM and PSNR over a 30-second sample. All three metrics are derived from the same pixel data but with different algorithms — if VMAF says "poor" but SSIM ≥ 0.98 or PSNR ≥ 40 dB, the encode is actually fine and VMAF is producing a measurement artefact (common on animation and other flat-coloured content, which is outside VMAF's training distribution). The log emits an explicit verdict line in that case: → SSIM/PSNR say the encode is actually fine; VMAF score is a measurement artefact. The VMAF-threshold rejection still fires on the raw VMAF score — the cross-check is purely diagnostic — so users debugging a rejection can open the job's ffmpeg output, see the numbers, and decide whether to lower the threshold for this kind of content.

Fixed

• VMAF analysis: bimodal scores on visually-identical encodes (e.g. sibling TV episodes scoring 49.5 and 96.3 at the same CQ and preset) traced to the filter-graph trim=start:duration approach picking different first frames in the reference vs encoded streams when the source had any of: VFR timestamps, non-zero container start_pts, interlaced fields, or keyframe-offset boundaries. Once the first frame misaligned, every subsequent pairwise compare was time-shifted and the score cratered. The 0.3.2 fix normalised pixel format and resolution but didn't touch temporal alignment, which is what was actually breaking. New behaviour:
  • For content ≤ 45 minutes (all TV episodes and most documentaries): VMAF compares the whole file — no trim, no sampling, no PTS math. The most reliable configuration possible.
  • For longer content (movies): sample a 90-second window at 33% into the file using input-level -ss (accurate seek in modern ffmpeg) applied identically to both inputs plus -t on the output — both streams are seeked to the same PTS before the filter graph sees them, so no in-filter trim is needed.
• History tab no longer labels VMAF-rejected jobs as "Converted (no savings)". A job whose encode scored below the minimum VMAF threshold is now recorded as Kept original — VMAF below threshold, which matches reality (the converted file was discarded and the source left in place). Encodes discarded because the output was larger than the source now read Kept original — encode was larger than source for the same reason. The plain Converted (no savings) label is reserved for real conversions that happened to break even on disk usage.

Docker images

Pinned to this release:

docker pull ghcr.io/i-ial9000/shrinkerr:0.3.3         # portable (amd64 + arm64, CPU)
docker pull ghcr.io/i-ial9000/shrinkerr:0.3.3-nvenc   # x86_64 + NVIDIA GPU (NVENC)
docker pull ghcr.io/i-ial9000/shrinkerr:0.3.3-edge    # as above but ffmpeg master
docker pull ghcr.io/i-ial9000/shrinkerr:0.3.3-edge-nvenc

Floating tags (:latest, :edge, :nvenc, :edge-nvenc) track the most recent release of each lineage.