Skip to content

docs(bench): v0.13 — per-shot CRF steering is an honest negative on the CRF axis #34

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lusoris merged 1 commit into from
Jun 20, 2026
Merged

docs(bench): v0.13 — per-shot CRF steering is an honest negative on the CRF axis #34

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
34 changes: 34 additions & 0 deletions docs/development/bench-results.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -484,6 +484,40 @@ incumbent `roi=1` banding mode already captures the banding win *without* the
texture-starving cost. A structure-tensor + CSF masking discriminator (new GPU
output) is the only path that could revive it — a documented follow-up, not taken.

## v0.13 — per-shot CRF steering (ADR-0132): honest negative on the CRF axis

Tested the per-shot complexity-budget lever as a **simple, hand-tuned** mapping:
pass 1 extracts per-frame complexity via the `lavfi.pelorus.*` metadata (ADR-0136)
and aggregates per shot; pass 2 biases x265 rate-control per shot via `--zones`
(more bits to complex shots, fewer to simple). Real ffmpeg + libx265, a 3-shot
BBB clip (per-shot complexity 0.065 / 0.052 / 0.032 — a real ~2× spread), RD curve
(CRF 26/30/34/38), BD-rate vs flat CRF.

First attempt used `zones=...,crf=` — **invalid**: x265 zones support only `q=`
(force QP) and `b=` (bitrate factor), not `crf=`, so it mis-parsed into a broken
half-bitrate encode. Corrected to `b=` factors (complex shot `b=1.16`, simple
`b=0.82`):

| metric | BD-rate | verdict |
|---|---|---|
| VMAF | **+8.81 %** | loss |
| SSIMULACRA2 | **+7.35 %** | loss |

Per-shot redistribution **loses** — it spent ~5 % more bits *and* scored slightly
lower at every CRF. Root cause (predicted from first principles): **CRF is already
a constant-quality target**, so x265 allocates more bits to complex shots on its
own; pushing *beyond* constant-quality toward complex shots moves bits off the
quality/bitrate convex hull. The per-shot win in the literature is for
**budget/VBR** allocation or **per-shot resolution**, not pure CRF.

This is the 4th measured negative in the "compete with the encoder's mature
rate-control" space (after fp16, subgroup, and the perceptual-AQ map, v0.11/v0.12)
— Pelorus's edge is **preprocessing the source**, not out-guessing CRF/2-pass/
mbtree. ADR-0132 stays *Proposed*; the lever now depends on the autotune
(ADR-0106) finding a mapping that beats flat CRF, against a low CRF-axis ceiling.
The durable result is the ADR-0136 metadata extraction path (shipped), which
makes this — and any future per-shot/autotune work — measurable from the shell.

## Open / next

1. **SVT-AV1 ROI on real content**: the v0.10 synthetic gain is modest; re-run
Expand Down